Este articulo muestra el poder fundamental de un cuidadoso trabajo de datos históricos a largo plazo. Con el poder de informar los debates políticos actuales.
https://t.co/JuOSVL35FJ
Abstract
I. Introduction
Aunque la riqueza inmobiliaria representa, por término medio, aproximadamente la mitad de la riqueza nacional en una economía típica (Piketty, 2014), se carece de datos sobre la rentabilidad total de la vivienda (revalorización de los precios más alquileres) (Shiller, 2000, ofrece algunos datos históricos sobre los precios de la vivienda, pero no sobre los alquileres). En este artículo nos basamos en un trabajo más exhaustivo sobre los precios de la vivienda (Knoll, Schularick y Steger, 2017) y en datos de nueva construcción sobre los alquileres (Knoll, 2017) para poder hacer un seguimiento de los rendimientos totales del mayor componente del capital social nacional.
También seguimos el trabajo anterior de documentar los rendimientos anuales de las acciones, bonos y letras, pero también en este caso hemos llevado el proyecto más lejos. Volvemos a calcular todas estas medidas a partir de fuentes originales, mejoramos los vínculos a través de algunas discontinuidades históricas importantes del mercado (por ejemplo, cierres de mercados y otras lagunas asociadas a guerras e inestabilidad política), y en varios casos accedemos a fuentes de datos brutos nuevas y no utilizadas anteriormente. En todos los casos, también hemos recurrido a fuentes auxiliares para validar externamente nuestros datos, y más de 100 páginas de material en línea documentan nuestras fuentes y métodos. Nuestro trabajo proporciona a los investigadores la primera amplia base de datos no comercial de rendimientos históricos de acciones, bonos y letras -y la única base de datos de rendimientos de la vivienda- con la cobertura más amplia tanto de países como de años.
La segunda vertiente conexa de la literatura es el análisis comparativo de los balances nacionales a lo largo del tiempo, como en Goldsmith (1985). Más recientemente, Piketty y Zucman (2014) han reunido datos de las cuentas nacionales y de otras fuentes que rastrean el desarrollo de la riqueza nacional durante largos periodos de tiempo. También calculan las tasas de rendimiento del capital dividiendo la renta agregada del capital en las cuentas nacionales por el valor agregado del capital, también procedente de las cuentas nacionales.
Nuestro trabajo es complementario y suplementario al suyo. Es complementario porque la perspectiva del precio de los activos y el enfoque de las cuentas nacionales están unidos en última instancia por normas e identidades contables. Utilizando valoraciones de mercado, podemos corroborar y mejorar las estimaciones de los rendimientos del capital que importan para la dinámica de la desigualdad de la riqueza. Nuestros datos de rentabilidad a largo plazo también complementan el trabajo de Piketty y Zucman (2014) en el sentido de que ampliamos en gran medida el número de países para los que podemos calcular las tasas reales de rentabilidad hasta finales del siglo XIX.
Los datos que reunimos pueden arrojar luz sobre debates de investigación activos que abarcan desde la fijación del precio de los activos hasta la desigualdad. Por ejemplo, en un área de investigación polémica, la acumulación de capital, la expansión de la participación del capital en la renta y la tasa de crecimiento de la economía en relación con la tasa de rendimiento del capital ocupan un lugar central en el actual debate suscitado por Piketty (2014) sobre la evolución de la riqueza, la renta y la desigualdad. ¿Qué dicen los patrones a largo plazo de las tasas de rendimiento de las diferentes clases de activos sobre estos posibles factores de desigualdad?
En muchas teorías financieras, las preferencias sobre el consumo actual frente al futuro, las actitudes hacia el riesgo y la covariación con el riesgo de consumo se manifiestan en las primas que las tasas de rendimiento de los activos de riesgo tienen sobre los activos seguros. Los rendimientos de las distintas clases de activos y sus correlaciones con el consumo constituyen el núcleo de la ecuación de Euler canónica del consumo que sustenta la teoría de la valoración de activos de los libros de texto (véase, por ejemplo, Mehra y Prescott 1985). Pero siguen existiendo tensiones entre la teoría y los datos, lo que lleva a seguir explorando nuevos paradigmas de valoración de activos, incluidas las finanzas conductuales. Nuestros nuevos datos añaden otra clase de activos de riesgo a la mezcla, la vivienda, y con ella nuevos retos.
En otra línea de investigación desencadenada por la crisis financiera, Summers (2014) trata de revivir la hipótesis del estancamiento secular planteada por primera vez en el discurso presidencial de la AEA de Alvin Hansen (1939). Las tendencias demográficas están empujando a las economías del mundo a un territorio desconocido, ya que el peso relativo de los prestatarios y los ahorradores está cambiando, y con ello aumenta la posibilidad de que el tipo de interés caiga en una cantidad insuficiente para equilibrar el ahorro y la inversión en el pleno empleo. ¿Qué pruebas hay de que esto sea así?
This article aims to bridge the gap between two related strands of the academic literature. The first strand is rooted in finance and is concerned with long-run returns on different assets. Dimson, Marsh, and Staunton (2009) probably marked the first comprehensive attempt to document and analyze long-run returns on investment for a broad cross-section of countries. Meanwhile, Homer and Sylla (2005) pioneered a multidecade project to document the history of interest rates.
The second related strand of literature is the analysis of comparative national balance sheets over time, as in Goldsmith (1985). More recently, Piketty and Zucman (2014) have brought together data from national accounts and other sources tracking the development of national wealth over long time periods. They also calculate rates of return on capital by dividing aggregate capital income in the national accounts by the aggregate value of capital, also from national accounts.
Our work is both complementary and supplementary to theirs. It is complementary because the asset price perspective and the national accounts approach are ultimately tied together by accounting rules and identities. Using market valuations, we are able to corroborate and improve the estimates of returns on capital that matter for wealth inequality dynamics. Our long-run return data are also supplementary to the work of Piketty and Zucman (2014) in the sense that we greatly extend the number of countries for which we can calculate real rates of return back to the late nineteenth century.
The evidence we gathered can shed light on active research debates reaching from asset pricing to inequality. For example, in one contentious area of research, the accumulation of capital, the expansion of capital’s share in income, and the growth rate of the economy relative to the rate of return on capital all feature centrally in the current debate sparked by Piketty (2014) on the evolution of wealth, income, and inequality. What do the long-run patterns on the rates of return on different asset classes have to say about these possible drivers of inequality?
In many financial theories, preferences over current versus future consumption, attitudes toward risk, and covariation with consumption risk all show up in the premiums that the rates of return on risky assets carry over safe assets. Returns on different asset classes and their correlations with consumption sit at the core of the canonical consumption Euler equation that underpins textbook asset pricing theory (see, e.g., Mehra and Prescott 1985). But tensions remain between theory and data, prompting further explorations of new asset pricing paradigms, including behavioral finance. Our new data add another risky asset class to the mix, housing, and with it come new challenges.
In another strand of research triggered by the financial crisis, Summers (2014) seeks to revive the secular stagnation hypothesis first advanced in Alvin Hansen’s (1939) AEA presidential address. Demographic trends are pushing the world’s economies into uncharted territory as the relative weight of borrowers and savers is changing, and with it the possibility increases that the interest rate will fall by an insufficient amount to balance saving and investment at full employment. What is the evidence that this is the case?
El hilo conductor de estos amplios temas de investigación es la noción de que la tasa de rendimiento es fundamental para comprender las fluctuaciones económicas a largo, medio y corto plazo. Pero, ¿qué tasa de rentabilidad? ¿Y cómo se mide? Para una determinada escasez de oferta de financiación, la tasa de riesgo es una medida de la rentabilidad de la inversión privada; en cambio, la tasa de seguridad desempeña un papel importante en la evaluación comparativa de la compensación por riesgo y suele estar vinculada a los debates sobre la configuración de la política monetaria y la noción de tasa natural. A continuación resumimos nuestras principales conclusiones.
I.A. Main Findings
We present four main findings:
1. On Risky Returns, rrisky
In terms of total returns, residential real estate and equities have shown very similar and high real total gains, on average about 7% a year. Housing outperformed equities before World War II. Since World War II, equities have outperformed housing on average but had much higher volatility and higher synchronicity with the business cycle. The observation that housing returns are similar to equity returns, but much less volatile, is puzzling. Like Shiller (2000), we find that long-run capital gains on housing are relatively low, around 1% a year in real terms, and considerably lower than capital gains in the stock market. However, the rental yield component is typically considerably higher and more stable than the dividend yield of equities so that total returns are of comparable magnitude.
Before World War II, the real returns on housing and equities (and safe assets) followed remarkably similar trajectories. After World War II this was no longer the case, and across countries equities then experienced more frequent and correlated booms and busts. The low covariance of equity and housing returns reveals that there could be significant aggregate diversification gains (i.e., for a representative agent) from holding the two asset classes.
2. On Safe Returns, rsafe
We find that the real safe asset return (bonds and bills) has been very volatile over the long run, more so than one might expect, and often even more volatile than real risky returns. Each world war was (unsurprisingly) a moment of very low safe rates, well below zero. So was the 1970s stagflation. The peaks in the real safe rate took place at the start of our sample, in the interwar period, and during the mid-1980s fight against inflation. In fact, the long decline observed in the past few decades is reminiscent of the secular decline that took place from 1870 to World War I. Viewed from a long-run perspective, the past decline and current low level of the real safe rate today is not unusual. The puzzle may well be why the safe rate was so high in the mid-1980s rather than why has it declined ever since.
Safe returns have been low on average in the full sample, falling in the 1%–3% range for most countries and peacetime periods. While this combination of low returns and high volatility has offered a relatively poor risk-return trade-off to investors, the low returns have also eased the pressure on government finances, in particular allowing for a rapid debt reduction in the aftermath of World War II.
3. On the Risk Premium, rrisky − rsafe
Over the very long run, the risk premium has been volatile. Our data uncover substantial swings in the risk premium at lower frequencies that sometimes endured for decades and far exceed the amplitudes of business-cycle swings.
In most peacetime eras, this premium has been stable at about 4%–5%. But risk premiums stayed curiously and persistently high from the 1950s to the 1970s, long after the conclusion of World War II. However, there is no visible long-run trend, and mean reversion appears strong. Interestingly, the bursts of the risk premium in the wartime and interwar years were mostly a phenomenon of collapsing safe returns rather than dramatic spikes in risky returns.
In fact, the risky return has often been smoother and more stable than the safe return, averaging about 6%–8% across all eras. Recently, with safe returns low and falling, the risk premium has widened due to a parallel but smaller decline in risky returns. But these shifts keep the two rates of return close to their normal historical range. Whether due to shifts in risk aversion or to other phenomena, the fact that safe returns seem to absorb almost all of these adjustments seems like a puzzle in need of further exploration and explanation.
4. On Returns Minus Growth, rwealth − g
Piketty (2014) argued that if investors’ return to wealth exceeded the rate of economic growth, rentiers would accumulate wealth at a faster rate and thus worsen wealth inequality. Using a measure of portfolio returns to compute “r minus g” in Piketty’s notation, we uncover an important finding. Even calculated from more granular asset price returns data, the same fact reported in Piketty (2014) holds true for more countries, more years, and more dramatically: namely, r ≫ g.
In fact, the only exceptions to that rule happen in the years in or around wartime. In peacetime, r has always been much greater than g. In the pre-World War II period, this gap was on average 5% (excluding World War I). As of today, this gap is still quite large, about 3%–4%, though it narrowed to 2% in the 1970s before widening in the years leading up to the global financial crisis.
One puzzle that emerges from our analysis is that while r minus g fluctuates over time, it does not seem to do so systematically with the growth rate of the economy. This feature of the data poses a conundrum for the battling views of factor income, distribution, and substitution in the ongoing debate (Rognlie 2015). The fact that returns to wealth have remained fairly high and stable while aggregate wealth increased rapidly since the 1970s suggests that capital accumulation may have contributed to the decline in the labor share of income over the recent decades (Karabarbounis and Neiman 2014). In thinking about inequality and several other characteristics of modern economies, the new data on the return to capital that we present here should spur further research.
II. A New Historical Global Returns Database
In this section, we discuss the main sources and definitions for the calculation of long-run returns. A major innovation is the inclusion of housing. Residential real estate is the main asset in most household portfolios, as we shall see, but so far very little has been known about long-run returns on housing. Our data on housing returns cover capital gains and imputed rents to owners and renters, the sum of the two being total returns.2 Equity return data for publicly traded equities are used, as is standard, as a proxy for aggregate business equity returns.3
The data include nominal and real returns on bills, bonds, equities, and residential real estate for Australia, Belgium, Denmark, Finland, France, Germany, Italy, Japan, the Netherlands, Norway, Portugal, Spain, Sweden, Switzerland, the United Kingdom, and the United States. The sample spans 1870 to 2015. Table I summarizes the data coverage by country and asset class.
Data Coverage
| Country | Bills | Bonds | Equity | Housing |
|---|---|---|---|---|
| Australia | 1870–2015 | 1900–2015 | 1870–2015 | 1901–2015 |
| Belgium | 1870–2015 | 1870–2015 | 1870–2015 | 1890–2015 |
| Denmark | 1875–2015 | 1870–2015 | 1873–2015 | 1876–2015 |
| Finland | 1870–2015 | 1870–2015 | 1896–2015 | 1920–2015 |
| France | 1870–2015 | 1870–2015 | 1870–2015 | 1871–2015 |
| Germany | 1870–2015 | 1870–2015 | 1870–2015 | 1871–2015 |
| Italy | 1870–2015 | 1870–2015 | 1870–2015 | 1928–2015 |
| Japan | 1876–2015 | 1881–2015 | 1886–2015 | 1931–2015 |
| Netherlands | 1870–2015 | 1870–2015 | 1900–2015 | 1871–2015 |
| Norway | 1870–2015 | 1870–2015 | 1881–2015 | 1871–2015 |
| Portugal | 1880–2015 | 1871–2015 | 1871–2015 | 1948–2015 |
| Spain | 1870–2015 | 1900–2015 | 1900–2015 | 1901–2015 |
| Sweden | 1870–2015 | 1871–2015 | 1871–2015 | 1883–2015 |
| Switzerland | 1870–2015 | 1900–2015 | 1900–2015 | 1902–2015 |
| United Kingdom | 1870–2015 | 1870–2015 | 1871–2015 | 1896–2015 |
| United States | 1870–2015 | 1871–2015 | 1872–2015 | 1891–2015 |
Like most of the literature, we examine returns to national aggregate holdings of each asset class. Theoretically, these are the returns that would accrue for the hypothetical representative-agent investor holding each country’s portfolio. An advantage of this approach is that it captures indirect holdings much better, although it leads to some double-counting, thereby boosting the share of financial assets over housing somewhat. The differences are described in Online Appendix O.4
II.A. The Composition of Wealth
Figure I shows the decomposition of economy-wide investable assets and capital stocks, based on data for five major economies at the end of 2015: France, Germany, Japan, the United Kingdom, and the United States.5 Investable assets shown in the left panel of Figure I (and in Online Appendix Table A.23) exclude assets that relate to intrafinancial holdings and cannot be held directly by investors, such as loans, derivatives (apart from employee stock options), financial institutions’ deposits, and insurance and pension claims. Other financial assets mainly consist of corporate bonds and asset-backed securities. Other nonfinancial assets are other buildings, machinery and equipment, agricultural land, and intangible capital. The capital stock is business capital plus housing. Other capital is mostly made up of intangible capital and agricultural land. Data are sourced from national accounts and national wealth estimates published by the countries’ central banks and statistical offices.6
Composition of Investable Assets and Capital Stock in the Major Economies
Color version of figure available online. Composition of total investable assets and capital stock. Average of the individual asset shares of France, Germany, Japan, the United Kingdom, and the United States, as of end-2015. Investable assets are defined as the gross total of economy-wide assets excluding loans, derivatives, financial institutions’ deposits, insurance, and pension claims. Other financial assets mainly consist of corporate bonds and asset-backed securities. Other nonfinancial assets are other buildings, machinery and equipment, agricultural land, and intangible capital. The capital stock is business capital plus housing. Other capital is mostly made up by intangible capital and agricultural land. Data are sourced from national accounts and national wealth estimates published by the countries’ central banks and statistical offices.
Housing, equity, bonds, and bills make up over half of all investable assets in the advanced economies today, and nearly two-thirds if deposits are included. The right panel of Figure I shows the decomposition of the capital stock into housing and various other nonfinancial assets. Housing is about one-half of the outstanding stock of capital. In fact, housing and equities alone represent over half of total assets in household balance sheets (see Online Appendix Figures A.5 and A.6).
The main asset categories outside the direct coverage of this study are commercial real estate, business assets, and agricultural land; corporate bonds; pension and insurance claims; and deposits. But most of these assets represent claims of, or are closely related to, assets that we do cover. For example, pension claims tend to be invested in stocks and bonds; listed equity is a levered claim on business assets of firms; land and commercial property prices tend to comove with residential property prices; and deposit rates are either included in or are very similar to our bill rate measure.7
Our data also exclude foreign assets. Even though the data on foreign asset holdings are relatively sparse, the evidence that we do have—presented in Online Appendix O.4—suggests that foreign assets have, through history, only accounted for a small share of aggregate wealth, and the return differentials between domestic and foreign asset holdings are, with few exceptions, not that large. Taken together, this means that our data set almost fully captures the various components of the return on overall household wealth.
II.B. Historical Returns Data
1. Bill Returns
The canonical risk-free rate is taken to be the yield on Treasury bills, that is, short-term, fixed-income government securities. The yield data come from the latest vintage of the long-run macrohistory database (Jordà, Schularick, and Taylor 2017).8 Whenever data on Treasury bill returns were unavailable, we relied on either money market rates or deposit rates of banks from Zimmermann (2017). Because short-term government debt was rarely used and issued in the earlier historical period, much of our bill rate data before the 1960s actually consist of deposit rates.9
2. Bond Returns
These are conventionally the total returns on long-term government bonds. Unlike earlier cross-country studies, we focus on the bonds listed and traded on local exchanges and denominated in local currency. This focus makes bond returns more comparable with the returns of bills, equities, and housing. Moreover, this results in a larger sample of bonds, and on bonds that are more likely to be held by the representative household in the respective country. For some countries and periods we have made use of listings on major global exchanges to fill gaps where domestic markets were thin or local exchange data were not available (for example, Australian bonds listed in New York or London). Throughout the sample we target a maturity of around 10 years. For the second half of the twentieth century, the maturity of government bonds is generally accurately defined. For the pre-World War II period we sometimes had to rely on data for perpetuals, that is, very long-term government securities (such as the British consol). Although as a convention we refer here to government bills and bonds as “safe” assets, both are naturally exposed to inflation and default risk, for example. In fact, real returns on these assets fluctuate substantially over time, as we shall see (specifically, Sections V and VI).
3. Equity Returns
These returns come from a broad range of sources, including articles in economic and financial history journals, yearbooks of statistical offices and central banks, stock exchange listings, newspapers, and company reports. Throughout most of the sample, we aim to rely on indices weighted by market capitalization of individual stocks and a stock selection that is representative of the entire stock market. For some historical time periods in individual countries, however, we also make use of indices weighted by company book capital, by stock market transactions, or weighted equally due to limited data availability.
4. Housing Returns
We combine the long-run house price series introduced by Knoll, Schularick, and Steger (2017) with a novel data set on rents drawn from the PhD thesis of Knoll (2017). For most countries, the rent series rely on the rent components of the cost of living of consumer price indices constructed by national statistical offices. We then combine them with information from other sources to create long-run series reaching back to the late nineteenth century. To proxy the total return on the residential housing stock, our returns include both rented housing and owner-occupied properties.10 Specifically, wherever possible we use house price and rental indices that include the prices of owner-occupied properties and the imputed rents on these houses. Imputed rents estimate the rent that an owner-occupied house would earn on the rental market, typically by using rents of similar houses that are rented. This means that, in principle, imputed rents are similar to market rents and are simply adjusted for the portfolio composition of owner-occupied as opposed to rented housing. Imputed rents, however, are not directly observed and hence are less precisely measured than market rents, and are typically not taxed.11 To the best of our knowledge, we are the first to calculate total returns to housing in the literature for as long and as comprehensive a cross section of economies as we report.
5. Composite Returns
We compute the rate of return on safe assets, risky assets, and aggregate wealth, as weighted averages of the individual asset returns. To obtain a representative return from the investor’s perspective, we use the outstanding stocks of the respective asset in a given country as weights. To this end, we make use of new data on equity market capitalization (from Kuvshinov and Zimmermann 2018) and housing wealth for each country and period in our sample and combine them with existing estimates of public debt stocks to obtain the weights for the individual assets. A graphical representation of these asset portfolios and further description of their construction is provided in Online Appendix O.3. Tables A.28 and A.29 present an overview of our four asset return series by country, their main characteristics and coverage. The article comes with an extensive data appendix that specifies the sources we consulted and discusses the construction of the series in greater detail (see the Online Data Appendix, Sections U, V, and W for housing returns, and Section X for equity and bond returns).
II.C. Calculating Returns
In addition to individual asset returns, we present a number of weighted “composite” returns aimed at capturing broader trends in risky and safe investments and the “overall return” or “return on wealth.” Online Appendix O.3 provides further details on the estimates of country asset portfolios from which we derive country-year specific weights.
II.D. Constructing Housing Returns Using the Rent-Price Approach
This section briefly describes our methodology to calculate total housing returns. We provide further details as needed in Section III.C and Online Appendix U. We construct estimates for total returns on housing using the rent-price approach. This approach starts from a benchmark rent-price ratio
estimated in a baseline year (t = 0). For this ratio we rely on net rental yields from the Investment Property Database (IPD).13 We can then construct a time series of returns by combining separate information from a country-specific house price index series and a country-specific rent index series. For these indices, we rely on prior work on housing prices (Knoll, Schularick, and Steger 2017) and new data on rents (Knoll 2017). This method assumes that the indices cover a representative portfolio of houses. Under this assumption, there is no need to correct for changes in the housing stock, and only information about the growth rates in prices and rents is necessary.
III. Rates of Return: Aggregate Trends
Our headline summary data appear in Table II and Figure II. The top panel of Table II shows the full sample (1870–2015) results, and the bottom panel shows results for the post-1950 sample. Note that here, and throughout the article, rates of return are always annualized. Units are always expressed in percent per year, for raw data and for means and standard deviations. All means are arithmetic means, except when specifically referred to as geometric means.15 Data are pooled and equally weighted, that is, they are raw rather than portfolio returns. We always include wars so that results are not polluted by bias from omitted disasters. We do, however, exclude hyperinflation years (but only a few) to focus on the underlying trends in returns and to avoid biases from serious measurement errors in hyperinflation years, arising from the impossible retrospective task of matching within-year timing of asset and CPI price level readings, which can create a spurious, massive under- or overstatement of returns in these episodes.16
Global Real Rates of Return
Color version available online. Arithmetic average real returns p.a., unweighted, 16 countries. Consistent coverage within each country: each country-year observation used to compute the average has data for all four asset returns.
Global Real Returns
| Real returns | Nominal returns | |||||||
|---|---|---|---|---|---|---|---|---|
| Bills | Bonds | Equity | Housing | Bills | Bonds | Equity | Housing | |
| Panel A: Full sample | ||||||||
| Mean return p.a. | 1.03 | 2.53 | 6.88 | 7.06 | 4.58 | 6.06 | 10.65 | 11.00 |
| Standard deviation | 6.00 | 10.69 | 21.79 | 9.93 | 3.32 | 8.88 | 22.55 | 10.64 |
| Geometric mean | 0.83 | 1.97 | 4.66 | 6.62 | 4.53 | 5.71 | 8.49 | 10.53 |
| Mean excess return p.a. | 1.51 | 5.85 | 6.03 | |||||
| Standard deviation | 8.36 | 21.27 | 9.80 | |||||
| Geometric mean | 1.18 | 3.77 | 5.60 | |||||
| Observations | 1,767 | 1,767 | 1,767 | 1,767 | 1,767 | 1,767 | 1,767 | 1,767 |
| Panel B: Post-1950 | ||||||||
| Mean return p.a. | 0.88 | 2.79 | 8.30 | 7.42 | 5.39 | 7.30 | 12.97 | 12.27 |
| Standard deviation | 3.42 | 9.94 | 24.21 | 8.87 | 4.03 | 9.81 | 25.03 | 10.14 |
| Geometric mean | 0.82 | 2.32 | 5.56 | 7.08 | 5.31 | 6.88 | 10.26 | 11.85 |
| Mean excess return p.a. | 1.91 | 7.42 | 6.54 | |||||
| Standard deviation | 9.21 | 23.78 | 9.17 | |||||
| Geometric mean | 1.51 | 4.79 | 6.18 | |||||
| Observations | 1,022 | 1,022 | 1,022 | 1,022 | 1,022 | 1,022 | 1,022 | 1,022 |
Notes. Annual global returns in 16 countries, equally weighted. Period coverage differs across countries. Consistent coverage within countries: each country-year observation used to compute the statistics in this table has data for all four asset returns. Excess returns are computed relative to bills.
The first key finding is that residential real estate, not equity, has been the best long-run investment over the course of modern history. Although returns on housing and equities are similar, the volatility of housing returns is substantially lower, as Table II shows. Returns on the two asset classes are in the same ballpark—around 7%—but the standard deviation of housing returns is substantially smaller than that of equities (10% for housing versus 22% for equities). Predictably, with thinner tails, the compounded return (using the geometric average) is vastly better for housing than for equities—6.6% for housing versus 4.7% for equities. This finding appears to contradict one of the basic tenets of modern valuation models: higher risks should come with higher rewards.
Differences in asset returns are not driven by unusual events in the early pre-World War II part of the sample. Table II, Panel B makes this point. Compared to the full sample results in the top panel, the same clear pattern emerges: stocks and real estate dominate in terms of returns. Moreover, average returns post–1950 are similar to those for the full sample even though the postwar subperiod excludes the devastating effects of the two world wars. Robustness checks are reported in Online Appendix Figures A.1, A.2, and A.3. Briefly, the observed patterns are not driven by the smaller European countries in our sample. Figure A.1 shows average real returns weighted by country-level real GDP, for the full sample and the post-1950 period. Compared to the unweighted averages, equity performs slightly better, but the returns on equity and housing remain very similar, and the returns and riskiness of all four asset classes are very close to the unweighted series in Table II.
The results could be biased due to the country composition of the sample at different dates given data availability. Online Appendix Figure A.2 plots the average returns for sample-consistent country groups, starting at benchmark years—the later the benchmark year, the more countries we can include. Again, the broad patterns discussed above are largely unaffected.
We also investigate whether the results are biased due to the world wars. Online Appendix Figure A.3 plots the average returns in this case. The main result remains largely unchanged. Online Appendix Table A.3 also considers the risky returns during wartime in more detail to assess the evidence for rare disasters in our sample. Returns during both wars were indeed low and often negative, although returns during World War II in a number of countries were relatively robust.
Finally, our aggregate return data take the perspective of a domestic investor in a representative country. Online Appendix Table A.14 instead takes the perspective of a global USD-investor and assesses the USD value of the corresponding returns. The magnitude and ranking of returns are similar to those reported in Table II, although the volatilities are substantially higher. This is to be expected given that the underlying asset volatility is compounded by the volatility in the exchange rate. We also find somewhat higher levels of USD returns, compared with those in local currency.
What comes next in our discussion of raw rates of return? We look more deeply at risky rates of return, and delve into their time trends and the decomposition of housing and equity returns into the capital gain and yield components in greater detail in Section IV. We do the same for safe returns in Section V. But first, to justify our estimates, because these are new data, we have to spend considerable time to explain our sources, methods, and calculations. We compare our data to other literature in Section III.A. We subject the equity returns and risk premium calculation to a variety of accuracy checks in Section III.B. We also subject the housing returns and risk premium calculation to a variety of accuracy checks in Section III.C. Section III.D discusses the comparability of the housing and equity return series. For the purposes of our article, these very lengthy next four subsections undertake the necessary due diligence and discuss the various quality and consistency checks we undertook to make our data a reliable source for future analysis—and only after that is done do we proceed with analysis and interpretation based on our data.
However, we caution that all these checks may be as exhausting as they are exhaustive and a time-constrained reader eager to confront our main findings may jump to the end of this section and resume where the analytical core of the article begins at the start of Section IV.
III.A. Comparison To Existing Literature
Earlier work on asset returns has mainly focused on equities and the corresponding risk premium over safe assets (bills or bonds), starting with Shiller’s analysis of historical U.S. data (Shiller 1981), later extended to cover post-1920 Sweden and the United Kingdom (Campbell 1999), and other advanced economies back to 1900 (Dimson, Marsh, and Staunton 2009) or back to 1870 (Barro and Ursúa 2008). The general consensus in this literature is that equities earn a large premium over safe assets. The cross-country estimates of this premium vary between 7% in Barro and Ursúa (2008) and 6% in Dimson, Marsh, and Staunton (2009) using arithmetic means. Campbell (1999) documents a 4.7% geometric mean return premium instead.
We find a similarly high, though smaller, equity premium using our somewhat larger and more consistent historical data set. Our estimate of the risk premium stands at 5.9% using arithmetic means and 3.8% using geometric means (see Table II). This is lower than the estimates by Campbell (1999) and Barro and Ursúa (2008). The average risk premium is similar to that found by Dimson, Marsh, and Staunton (2009), but our returns tend to be slightly lower for the overlapping time period.17 Details aside, our data confirm the central finding of the literature on equity market returns: stocks earn a large premium over safe assets.
Studies on historical housing returns, starting with the seminal work of Robert Shiller (see Shiller 2000 for a summary), have largely focused on capital gains. The rental yield component has received relatively little attention and in many cases is missing entirely. Most of the literature pre-dating our work has therefore lacked the necessary data to calculate, infer, or discuss the total rates of return on housing over the long run. The few studies that take rents into account generally focus on the post-1970s U.S. sample and often on commercial real estate. Most existing evidence either places the real estate risk premium between equities and bonds or finds that it is similar to that for equities (see Ruff 2007; Francis and Ibbotson 2009; Ilmanen 2011; Favilukis, Ludvigson, and Van Nieuwerburgh 2017). Some studies have even found that over the recent period, real estate seems to outperform equities in risk-adjusted terms (Shilling 2003; Cheng, Lin, and Liu 2008).
The stylized fact from the studies on long-run housing capital appreciation is that over long horizons, house prices only grow a little faster than the consumer price index. But again, note that this is only the capital gain component in equation (1). Low levels of real capital gains to housing was shown by Shiller (2000) for the United States, and is also true, although to a lesser extent, in other countries, as documented in Knoll, Schularick, and Steger (2017). Our long-run average capital appreciation data for the United States largely come from Shiller (2000), with two exceptions. For the 1930s, we use the more representative index of Fishback and Kollmann (2015) that documents a larger fall in prices during the Great Depression. From 1975 onward, we use a Federal Housing Finance Agency index, which has a slightly broader coverage. Online Appendix M compares our series with Shiller’s data and finds that switching to Shiller’s index has no effect on our results for the United States. See also the online appendix of Knoll, Schularick, and Steger (2017) for further discussion.
However, our article turns this notion of low housing returns on its head—because we show that including the yield component in equation (1) in the housing return calculation generates a housing risk premium roughly as large as the equity risk premium. Prior to our work on historical rental yields, this finding could not be known. Coincidentally, in our long-run data we find that most of the real equity return also comes from the dividend yield rather than from real equity capital gains, which are low, especially before the 1980s. Thus the post-1980 observation of large capital gain components in equity and housing total returns is completely unrepresentative of the normal long-run patterns in the data, another fact underappreciated before now.
Data on historical returns for all major asset classes allow us to compute the return on aggregate wealth (see equation (7)). In turn, this return can be decomposed into various components by asset class, and into capital gains and yields, to better understand the drivers of aggregate wealth fluctuations. This connects our study to the literature on capital income and the stock of capital (or wealth) from a national accounts perspective. Even though national accounts and national balance sheet estimates have existed for some time (see Kuznets 1941; Goldsmith 1985), only recently have scholars systematized and compared these data to calculate a series of returns on national wealth.18
Piketty, Saez, and Zucman (2018) compute balance sheet returns on aggregate wealth and for individual asset classes using post-1913 U.S. data. Balance sheet return data outside the United States are sparse, although Piketty and Zucman (2014) provide historical estimates at benchmark years for three more countries and, after 1970, continuous data for an additional five countries. Online Appendix R compares our market-based return estimates for the United States with those of Piketty, Saez, and Zucman (2018). Housing returns are very similar. However, equity returns are several percentage points above our estimates and those in the market-based returns literature more generally. Part of this difference reflects the fact that balance sheet returns are computed to measure income before corporate taxes, whereas our returns take the household perspective and are therefore net of corporate tax. Another explanation for the difference is likely to come from some measurement error in the national accounts data.19 When it comes to housing, our rental yield estimates are broadly comparable and similar to those derived using the balance sheet approach, for a broad selection of countries and historical time spans.20
Our data set complements the market-based return literature by increasing the coverage in terms of assets, return components, countries, and years; improving data consistency and documentation; and making the data set freely available for future research. This comprehensive coverage can help connect the market-based return estimates to those centered around national accounts concepts. We hope that eventually, this can improve the consistency and quality of market-based returns and national accounts data.
III.B. Accuracy of Equity Returns
The literature on equity returns has highlighted two main sources of bias in the data: weighting and sample selection. Weighting biases arise when the stock portfolio weights for the index do not correspond with those of a representative investor or a representative agent in the economy. Selection biases arise when the selection of stocks does not correspond to the portfolio of the representative investor or agent. This second category also includes issues of survivorship bias and missing data bias arising from stock exchange closures and restrictions.
We consider how these biases affect our equity return estimates in this section. An accompanying Online Appendix Table A.29 summarizes the construction of the equity index for each country and time period, with further details provided in Online Appendix X.
1. Weighting Bias
The best practice when weighting equity indices is to use market capitalization of individual stocks. This approach most closely mirrors the composition of a hypothetical representative investor’s portfolio. Equally-weighted indices are likely to overweight smaller firms, which tend to carry higher returns and higher volatility.
The existing evidence from historical returns on the Brussels and Paris stock exchanges suggests that using equally-weighted indices biases returns up by around 0.5 percentage points, and their standard deviation up by 2–3 percentage points (Le Bris and Hautcoeur 2010; Annaert et al. 2011). The size of the bias, however, is likely to vary across markets and time periods. For example, Grossman (2017) shows that the market-weighted portfolio of UK stocks outperformed its equally-weighted counterpart over the period 1869–1929.
To minimize this bias, we use market-capitalization-weighted indices for the vast majority of our sample (see Online Appendix Table A.29 and Online Appendix X). Where market capitalization weighting was not available, we have generally used alternative weights such as book capital or transaction volumes, rather than equally-weighted averages. For the few equally-weighted indices that remain in our sample, the overall impact on aggregate return estimates ought to be negligible.
2. Selection and Survivorship Bias
Relying on an index whose selection does not mirror the representative investor’s portfolio carries two main dangers. First, a small sample may be unrepresentative of overall stock market returns. Second, a sample that is selected ad hoc, and especially ex post, is likely to focus on surviving firms, or successful firms, thus overstating investment returns. This second bias extends not only to stock prices but also to dividend payments, as some historical studies only consider dividend-paying firms.21 The magnitude of survivorship bias has generally been found to be around 0.5 to 1 percentage point (Nielsen and Risager 2001; Annaert, Buelens, and De Ceuster 2012), but in some time periods and markets it could be larger (see Le Bris and Hautcoeur 2010 for France).
As a first best, we always strive to use all-share indices that avoid survivor and selection biases. For some countries and time periods where no such indices were previously available, we have constructed new weighted all-share indices from original historical sources (e.g., early historical data for Norway and Spain). Where an all-share index was not available or newly constructed, we have generally relied on blue-chip stock market indices. These are based on an ex ante value-weighted sample of the largest firms on the market. It is updated each year and tends to capture the lion’s share of total market capitalization. Because the sample is selected ex ante, it avoids ex post selection and survivorship biases. Because historical equity markets have tended to be quite concentrated, blue-chip indices have been shown to be a good proxy for all-share returns (see Annaert et al. 2011). Finally, we include non-dividend-paying firms in the dividend yield calculation.
3. Stock Market Closures and Trading Restrictions
A more subtle form of selection bias arises when the stock market is closed and no market price data are available. One way of dealing with closures is to simply exclude them from the baseline return comparisons. But this implicitly assumes that the data are “missing at random”—that is, that stock market closures are unrelated to underlying equity returns. Existing research on rare disasters and equity premiums shows that this is unlikely to be true (Nakamura et al. 2013). Stock markets tend to be closed precisely when we would expect returns to be low, such as periods of war and civil unrest. Return estimates that exclude such rare disasters from the data will thus overstate stock returns.
To guard against this bias, we include return estimates for the periods of stock market closure in our sample. Where possible, we rely on alternative data sources to fill the gap, such as listings of other exchanges and over-the-counter transactions—for example, in the case of World War I Germany we use the over-the-counter index from Ronge (2002) and for World War II France we use the newspaper index from Le Bris and Hautcoeur (2010). In cases where alternative data are not available, we interpolate the prices of securities listed before and after the exchange was closed to estimate the return (if no dividend data are available, we also assume no dividends were paid).22
Even though this approach only gives us a rough proxy of returns, it is certainly better than excluding these periods, which effectively assumes that the return during stock market closures is the same as when the stock markets are open. In the end, we only have one instance of stock market closure for which we are unable to estimate returns—that of the Tokyo stock exchange in 1946–1947. Online Appendix H further assesses the impact of return interpolation on the key moments of our data and finds that over the full sample, it is negligible.
Table III shows the estimated stock returns during the periods of stock exchange closure in our sample. The first two columns show average and cumulative real returns, and the third and fourth columns show the nominal returns. Aside from the case of World War I Germany, returns are calculated by comparing the prices of shares listed before and after the market closure. Such a calculation may overstate returns because it selects only those companies that “survived” the closure. As an additional check, the last two columns of Table III show the inflation-adjusted change in market capitalization of stocks before and after the exchange was closed. This serves as a lower bound for investor returns because it would be as if we assumed that all delisted stocks went bankrupt (i.e., to a zero price) during the market closure.
Geometric Annual Average and Cumulative Total Equity Returns in Periods of Stock Market Closure
| Real returns | Nominal returns | Real capitalization | ||||
|---|---|---|---|---|---|---|
| Geometric | Geometric | Geometric | ||||
| Episode | average | Cumulative | average | Cumulative | average | Cumulative |
| Spanish Civil War, 1936–40 | −4.01 | −15.09 | 9.03 | 41.32 | −10.22 | −35.04 |
| Portuguese revolution, 1974–77 | −54.98 | −90.88 | −44.23 | −82.65 | −75.29 | −98.49 |
| Germany World War I, 1914–18 | −21.67 | −62.35 | 3.49 | 14.72 | ||
| Switzerland World War I, 1914–16 | −7.53 | −14.50 | −0.84 | −1.67 | −8.54 | −16.34 |
| Netherlands World War II, 1944–46 | −12.77 | −20.39 | −5.09 | −8.36 | ||
Notes. Cumulative and geometric average returns during periods of stock market closure. Estimated by interpolating returns of shares listed both before and after the exchange was closed. The change in market capitalization compares the capitalization of all firms before the market was closed, and once it was opened, and thus includes the effect of any new listings, delistings, and bankruptcies that occured during the closure.
Indeed, the hypothetical investor returns during the periods of market closure are substantially below market averages. In line with Nakamura et al. (2013), we label these periods “rare disasters.” The average per year geometric mean return ranges from a modestly negative –4% p.a. during the Spanish Civil War, to losses of roughly 55% p.a. during the three years after the Portuguese Carnation Revolution. Accounting for returns of delisted firms is likely to bring these estimates down further, as evinced by the virtual disappearance of the Portuguese stock market in the aftermath of the revolution.
Having said this, the impact of these rare events on the average cross-country returns (shown in Table II) is small, around –0.1 percentage points, precisely because protracted stock market closures are very infrequent. The impact on country-level average returns is sizable for Portugal and Germany (around –1 percentage point), but small for the other countries (–0.1 to –0.4 percentage points). Online Appendix G provides a more detailed analysis of returns during consumption disasters. On average, equity returns during these times are low, with an average cumulative real equity return drop of 6.7% during the disaster years.
Last, Nakamura et al. (2013) also highlight a more subtle bias arising from asset price controls. This generally involves measures by the government to directly control transaction prices, as in Germany during 1943–47, or to influence the funds invested in the domestic stock market (and hence the prices) via controls on spending and investment, as in France during World War II (Le Bris 2012). These measures are more likely to affect the timing of returns rather than their long-run average level and should have little impact on our headline estimates. For example, Germany experienced negative nominal and real returns despite the World War II stock price controls; and even though the policies it enacted in occupied France succeeded in generating high nominal stock returns, the real return on French stocks during 1940–44 was close to zero. Both instances were followed by sharp drops in stock prices when the controls were lifted.23
III.C. Accuracy of Housing Returns
The biases that affect equity returns—weighting and selection—can also apply to returns on housing. There are also other biases that are specific to housing return estimates. These include costs of running a housing investment and the benchmarking of rent-price ratios to construct the historical rental yield series. We discuss these problematic issues in this section. Our focus is mainly on rental yield data, as the accuracy and robustness of the house price series has been extensively discussed in Knoll, Schularick, and Steger (2017) in their online appendix.
1. Maintenance Costs
Any homeowner incurs costs for maintenance and repairs, which lower the rental yield and thus the effective return on housing. We deal with this issue by the choice of the benchmark rent-price ratios. Specifically, we anchor to the Investment Property Database (IPD), whose rental yields reflect net income—net of property management costs, ground rent, and other unrecoverable expenditure—as a percentage of the capital employed. The rental yields calculated using the rent-price approach detailed in Section II.D are therefore net yields. To enable a like-for-like comparison, our historical benchmark yields are calculated net of estimated running costs and depreciation. Running costs are broadly defined as housing-related expenses excluding interest, taxes, and utilities—that is, maintenance costs, management, and insurance fees.
Applying the rent-price approach to net yield benchmarks assumes that running costs remain stable relative to gross rental income over time within each country. To check this, Figure III presents historical estimates of running costs and depreciation for Australia, France, the United Kingdom, and the United States, calculated as the sum of the corresponding housing expenditures and fixed capital consumption items in the national accounts. The left panel presents these as a proportion of total housing value and the right panel as a proportion of gross rent. Relative to housing value, costs have been stable over the past 40 years, but were somewhat higher in the early to mid-twentieth century. This is to be expected because these costs are largely related to structures, not land, and structures constituted a greater share of housing value in the early twentieth century (Knoll, Schularick, and Steger 2017). In addition, structures themselves may have been of poorer quality in past times. When taken as a proportion of gross rent, however, as shown in the right panel of Figure III, housing costs have been relatively stable, or at least not higher historically than they are today. This is likely because gross yields and costs are low today, whereas historically yields and costs were higher, with the two effects more or less canceling out. This suggests that the historical rental yields that we have calculated using the rent-price approach are a good proxy for net yields.
Costs of Running a Housing Investment
Color version available online. Total costs include depreciation and all other housing-related expenses excluding interest, taxes, and utilities (mainly maintenance and insurance payments). Costs are estimated as the household consumption of the relevant intermediate housing input, or fixed housing capital, in proportion to total housing wealth (left), or total gross rent (right).
2. Rental Yield Benchmarking
To construct historical rental yield series using the rent-price approach, we start with a benchmark rent-price ratio from the IPD and extend the series back using the historical rent and house price indices (see Section II.D).24 This naturally implies that the level of returns is sensitive to the choice of the benchmark ratio. Moreover, past errors in rent and house price indices can potentially accumulate over time and may cause one to substantially over- or understate historical rental yields and housing returns. If the historical capital gains are overstated, the historical rental yields will be overstated, too.
To try to avert such problems, we corroborate our rental yield estimates using a wide range of alternative historical and current-day sources. The main source of these independent comparisons comes from estimates using the balance sheet approach and national accounts data. As shown in equation (12), the “balance sheet” rental yield is the ratio of nationwide net rental income to total housing wealth. Net rental income is computed as gross rents paid less depreciation, maintenance, and other housing-related expenses (excluding taxes and interest), with all data taken from the national accounts. The balance sheet approach gives us a rich set of alternative rental yield estimates both for the present day and even going back in time to the beginning of our sample in a number of countries. The second source for historical comparisons comes from advertisements in contemporary newspapers and various other contemporary publications. Third, we also make use of alternative current-day benchmarks based on transaction-level market rent data, and the rental expenditure and house price data from Numbeo.com.25 For all these measures, we adjust gross yields down to obtain a proxy for net rental yields.
Historical sources offer point-in-time estimates that avoid the cumulation of errors but can nevertheless be imprecise. The balance sheet approach relies on housing wealth and rental income data, both of which are subject to potential measurement error. For housing wealth, it is inherently difficult to measure the precise value of all dwellings in the economy. Rental income is largely driven by the imputed rents of homeowners, which have to be estimated from market rents by matching the market rent to owner-occupied properties based on various property characteristics. This procedure can suffer from errors in the survey data on property characteristics and market rents, and the matching algorithm.26 Newspaper advertisements are tied to a specific location and are often biased toward cities. And transaction-level or survey data sometimes only cover the rental sector, rather than both renters and homeowners.
Given the potential measurement error in all the series, our final rental yield series uses data from both the rent-price approach and the alternative benchmarks listed above. More precisely, we use the following method to construct our final “best-practice” rental yield series. If the rent-price approach estimates are close to alternative measures, we keep the rent-price approach data. This is the case for most historical periods in our sample. If there is a persistent level difference between the rent-price approach and alternative estimates, we adjust the benchmark yield to better match the historical and current data across the range of sources. This is the case for Australia and Belgium. If the levels are close for recent data but diverge historically, we adjust the historical estimates to match the alternative benchmarks. For most countries, such adjustments are small or only apply to a short time span, but for Finland and Spain they are more substantial. Online Appendix U details the alternative sources and rental yield construction, including any such adjustments, for each country.
How large is the room for error in our final housing return series? To get a sense of the differences, Figure IV compares the rent-price approach of net rental yield estimates (black diamonds) with those using the balance sheet approach (brown triangles; see color version of figure online). The first three panels show the time series of the two measures for France, Sweden, and the United States, and the bottom right panel shows the correlation between changes in rent-price and balance sheet yields in nine countries (Australia, Denmark, France, Germany, Italy, Japan, Sweden, the United Kingdom, and the United States).27 The level of the rent-price ratio using the two approaches is similar in the modern day and historically.28 The two yield measures follow a very similar time series pattern.
Comparison of the Rent-Price and Balance-Sheet Approaches for Historical Rental Yields
The rent-price approach uses the baseline estimates in this article. The balance sheet approach estimates the net yield in each year as total rental expenditure less housing running costs and depreciation, in proportion to total housing wealth.
Table IV provides a more comprehensive comparison. The first two columns present the arithmetic and geometric mean, and the standard deviation, for the baseline measures of equity and housing annual real total returns in our sample (also shown in Table II). The third column instead uses the lowest possible initial benchmark for the housing series.29 The resulting returns are around 0.8 percentage points lower in both arithmetic and geometric mean terms. The fourth column instead uses the highest available benchmark, thus raising housing returns by 0.8 percentage points. The fifth column uses historical benchmarks for all rental yield series before 1980, that is, we use these benchmarks as the main source for the yields, and only use the rent-price approach for interpolation.30 This makes very little difference to the returns, lowering them by around 0.2 percentage points. The last column uses the balance sheet approach as the baseline estimate for the current and historical period. It then uses the rent-price approach to fill the gaps and interpolate between the balance sheet estimates.31 Finally, we compute the total balance sheet return on housing as the sum of capital gains and the balance sheet yield.32 The resulting return is 0.8 percentage points lower than our baseline estimates.
Impact of Using Different Rental Yield Benchmarks
| Equity | Housing | |||||
|---|---|---|---|---|---|---|
| Low | High | Balance | ||||
| initial | initial | Historical | sheet | |||
| Baseline | benchmark | benchmark | benchmarks | approach | ||
| Mean return p.a. | 6.88 | 7.06 | 6.29 | 7.89 | 6.83 | 6.30 |
| Standard deviation | 21.79 | 9.93 | 9.89 | 10.03 | 9.93 | 9.95 |
| Geometric mean | 4.66 | 6.62 | 5.85 | 7.45 | 6.39 | 5.86 |
| Observations | 1,767 | 1,767 | 1,767 | 1,767 | 1,767 | 1,767 |
Note. Average total real returns across 16 countries, equally weighted.
Taken together, this analysis suggests that the potential margins for error are small. Even under the more stringent checks, housing returns remain within a percentage point of our baseline estimates. The average return is always similar to equities in arithmetic mean terms and always above equities when using the geometric mean.
3. Geographic Coverage and Selection Biases
Our data aim to approximate the return on a representative agent’s housing portfolio. Selection bias means that the selection of properties in our data set does not mirror the balance sheet of the representative agent. The main reason for this bias is selective geographical coverage. Housing returns can vary a lot by location, and our data are based on a sample of housing transactions.
To make our samples as representative as possible, we strive to attain a broad geographic coverage for both house price and rental data. Knoll, Schularick, and Steger (2017) discuss the potential location biases in house price data but find that the house price trends in their, and hence our, data set should be representative of country-level trends. When it comes to rents, the benchmark IPD yields are based on portfolios of institutional investors, which are slightly biased toward cities. This would lead to lower yields than the national average. On the other hand, investors may select higher-yielding properties within any given city. Comparisons with aggregate balance sheet approach data and alternative estimates indicate that, on average, IPD yields tend to be representative at country level. Furthermore, IPD yields are capitalization weighted, which again better captures the yield on a representative portfolio. Finally, we aim for national coverage with the historical rental indices used for extrapolation and historical balance sheet benchmarks.
Despite this, our house price and rental data are likely somewhat biased toward cities and urban areas, especially for historical periods—simply because urban housing data are more widely available and researched. Even though this would affect the level of capital gain and yield, it should have little influence on total returns, since cities tend to have higher capital gains but lower rental yields.33 In addition, Knoll, Schularick, and Steger (2017) show that the rural-urban divide has a relatively small impact on capital gains. Relatedly, we can establish some bounds on how much our rental yields can vary with the choice of location. In 2013, Numbeo.com data suggest that price-rent ratios in and out of city centers differ by less than three times annual rent. The rental yield is the inverse of these price-rent ratios. This motivates us to construct a lower-bound rent-price ratio as
and an upper-bound rent-price ratio asfor each country in 2013 to estimate upper and lower bounds of our housing returns depending on the choice of location. Given the currently high price-rent ratios, these adjustments have a relatively small impact on our data. Figure V shows that increasing or reducing the price-rent ratio by three changes annual return estimates by about ±1 percentage point a year relative to our preferred baseline.
Sensitivity of Housing Returns to a Rent-Price Location Correction
Bars show the arithmetic- and geometric-average housing returns for selected subperiods. Error bars show the impact on historical returns of increasing or reducing the benchmark price/rent ratio by ±3, which broadly captures the difference between in- and out-of-city-center locations.
This suggests that the level of housing returns in our data set should be representative of a country-wide portfolio. Still, it could be that returns on locations not included in our sample display higher volatility. For example, the post-1975 U.S. indices are based on conforming mortgages and may exclude the more volatile segments of the market. To assess the likely magnitude of this bias, Table V compares the recent level and volatility of the U.S. conforming mortgage based OFHEO house price indices with those that cover other segments of the market as well, which are sourced from Zillow.34 Comparing the second and third columns of Table V, the nationwide moments of the data are similar across the two measures—but as expected, the OFHEO data display slightly higher real capital gains and slightly lower volatility, because they have a less comprehensive coverage of the areas that were hit hardest by the subprime crisis, which receives a relatively high weight in the 1995–2015 U.S. sample used here.
Level and Volatility of Real Housing Capital Gains at Different Levels of Coverage and Aggregation
| Baseline | Zillow | ||||
|---|---|---|---|---|---|
| National | National | State | County | ZIP code | |
| Mean real capital gain p.a. | 1.42 | 0.79 | 1.07 | 0.53 | 0.92 |
| Standard deviation | 4.67 | 5.67 | 6.05 | 6.28 | 7.46 |
Notes. U.S. data, 1995–2015. Average annual real capital gain and standard deviation of house prices. Baseline data are sourced from the OFHEO index. Zillow data are sourced from the Zillow Home Value Index, which covers around 95% of the U.S. housing stock, and are averages of monthly values. National data are the returns and volatility of prices for a nationwide housing index, and the other figures cover a representative state, county, or ZIP code level portfolio, respectively.
Table V also shows that the volatility of the housing series increases as we move from the aggregate portfolio (second column) to the subnational and local level. The standard deviation of ZIP code–level housing returns is roughly one-third higher than that in the national data. If investors owned one undiversified house whose price tracks the neighborhood index, the risk and return characteristics of this portfolio would be somewhat closer to those of the aggregate equity index, although the gap would still be large.
Of course, it is much more difficult to invest in a diversified housing portfolio than a well diversified equity portfolio. That being said, Benhabib and Bisin (2016) show that most equity is also held in an undiversified manner. The data regarding returns on individual housing and private equity returns are, however, very sparse at this point in time. To understand exactly how these risk-return characteristics play out at a disaggregated level, a more detailed study of individual portfolios and returns is necessary. This would be a worthy goal of future research.
Another selection bias comes about from the fact that rent data largely come from the rental market, whereas the majority of housing stock is held by owner-occupiers. To guard against this, we rely on rental indices that, whenever possible, account for the growth of imputed rents. We also benchmark our series to the balance sheet yield data constructed to cover the total housing stock. Still, imputed rents are measured with error and may not reflect the cost that the homeowner would pay on the rental market. If owning is relatively cheaper than renting—for example, due to tax exemptions or long-run house price appreciation—homeowners would purchase larger or better houses than they would rent, and imputed rents would overstate the value of housing services accruing to homeowners. On the other hand, buying a house is subject to credit constraints, which means that renters can afford better houses than homeowners, all else equal. In this case, imputed rents would understate the flow value of housing services. Overall, the direction of any potential bias is unclear and leaves much scope for future work.
Finally, the portfolio selection in the price and rent series is subject to survivorship bias. In general, our price and rental yield estimates aim to capture transaction or appraisal values, and rental costs on a broad and impartially selected portfolio of properties. Some survivorship bias may enter the series for the following reasons. First, indices that rely on an ex post selection of cities may inadvertently choose the more “successful” cities over the less successful ones. Second, houses that decline in value are likely to lose liquidity and be sold less frequently, hence carrying a lower weight in the index. Third, chain-linking historical house price and rent indices to compute annual returns will generally ignore the impact of large destructions of the housing stock, particularly those occurring during wartime.
Several factors suggest that the impact of survivorship bias on housing returns should be limited. First, Figure V and Knoll, Schularick, and Steger (2017) show that any location-specific bias in our estimates is likely to be small. Second, if the magnitude of survivorship bias is similar to that in equity markets (Section III.B), the bias is also unlikely to be large. Third, the low liquidity and weight of houses with declining prices is in some ways similar to the documented negative returns on delisted equities (Shumway 1997; Shumway and Warther 1999) which, generally, cannot be incorporated into the stock return series due to the lack of data. Therefore this bias should be less of a concern when comparing housing and equity returns. Finally, similarly to the stock market closures discussed in Section III.B, even though capital stock destruction during wars can have a substantial impact on returns in specific years, it is unlikely to profoundly affect cross-country long-run returns due to the rarity of such events.35 As Figure IX shows later, the main facts in the data are similar for countries that experienced major war destruction on their own territory versus countries that did not (e.g., Australia, Canada, Sweden, Switzerland, and the United States). Online Appendix Table A.5 shows that housing offers a similar return relative to equity on average even after wars are excluded.
4. Returns on Real Estate Investment Trusts
Another way to check our housing returns is to compare them to the historical returns on housing investment trusts. These trusts offer independent estimates of returns. Real estate investment trusts (REITs), are investment funds that specialize in the purchase and management of residential and commercial real estate. Many of these funds list their shares on the local stock exchange. The return on these shares should closely track total real estate returns. Differences will arise because the REIT portfolio is more geographically concentrated, its assets often contain nonresidential property, and share price fluctuations may reflect expectations of future earnings and sentiment, as well as underlying portfolio returns. Furthermore, the REIT portfolio returns should be net of taxes and transaction costs as well as housing running costs, and may thus be somewhat lower than our housing series. Still, returns on the REIT portfolio should be comparable to housing and can be used to check the general plausibility of our return series.
Figure VI compares our historical housing returns (dashed line) with those on investments in REITs (solid line) in France and the United States, two countries for which longer-run REIT return data are available. The REIT returns series for France refers to shares of the fund La Fourmi Immobilière (see Simonnet, Gallais-Hamonno, and Arbulu 1998). The fund acquired a portfolio of 15 properties in Paris between 1900 and 1913, worth around €36 million at 2015 prices, and its shares were listed on the Paris stock exchange between 1904 and 1997. We exclude the period after 1985, when La Fourmi Immobilière was taken over by AGF. For the United States, we use the FTSE NAREIT residential total return index after 1994, and the general FTSE equity NAREIT before. REIT returns have to be unlevered to capture the returns on the REIT housing portfolio. La Fourmi Immobilière had an unlevered balance sheet structure, hence we do not adjust their returns. We assume a REIT leverage of 45% for the United States following Giacomini, Ling, and Naranjo (2015). Returns for France are presented as decadal moving averages, and for the United States as five-year moving averages, given the shorter span of the U.S. data.
Returns on Housing Compared to Real Estate Investment Funds
Total real return on housing, and shares of housing investment firms in France and the U.S. Moving averages. Following Giacomini, Ling, and Naranjo (2015), we assume a 45% leverage ratio for US REITs.
Comparing the solid and dashed lines in Figure VI, we can see that the long-run levels of unlevered REIT and housing returns are remarkably similar. The time trend also follows a similar pattern, especially in France. The REIT returns, however, tend to be somewhat more volatile—most likely because they reflect changes in the market’s valuations of future earnings and the current portfolio performance. The REIT returns also seem to be affected by the general ups and downs of the stock market: for example, the 1987 Black Monday crash and dot-com bust in the United States, as well as the 1930s Great Depression and 1960s stock crises in France. This suggests that the valuations of the funds’ housing portfolios may have been affected by shifts in general stock market sentiment, possibly unrelated to housing market fundamentals.
Overall, the returns on real estate investment funds serve to confirm the general housing return level in our data set. The comparison also suggests that returns in housing markets tend to be smoother than those in stock markets. The next section examines various factors that can affect the comparability of housing and equity returns more generally.
III.D. Comparability of Housing and Equity Returns
Even if the fundamentals driving housing and equity returns (expected dividend/profit, and rental flows) are similar, investor returns for the two asset classes may differ for a number of reasons including taxes, transaction costs, and the financial structure of the investment claim. In this subsection we consider such comparability issues.
1. Transaction Costs
The conventional wisdom is that while bonds and equities can be purchased with low transaction costs and at short notice, the seller of a house typically incurs significant costs. We provide a rough estimate of how transaction costs affect our return estimates for housing. We perform a simple back-of-the-envelope calculation to do this using contemporary data on average holding periods of residential real estate and average transaction costs incurred by the buyer. According to the OECD (2012), average round-trip transaction costs across 13 of the 16 countries in our sample amount to about 7.7% of the property’s value.36
However, these simple cost ratios need to be adjusted for the typical trading frequency of each asset. According to the American Community Survey of 2007, more than 50% of U.S. homeowners had lived in their current home for more than 10 years. Current average holding periods are similar in, for example, the United Kingdom, Australia, and the Netherlands. Another way to estimate housing turnover is using housing sales data, which for the United States gives us an average holding period of close to 20 years.37 Either way, accounting for transaction costs would thus lower the average annual return to housing by less than 100 basis points (e.g., 77 basis points a year based on a 7.7% cost incurred every 10 years).
For equities, the cost of each individual transaction is much smaller, but the number of transactions is much higher. Jones (2002) estimates that at the New York Stock Exchange over the period 1900–2001, the average transaction cost was around 80 bps (half bid-ask spread of 30 bps plus commission rate of 50 bps), and turnover was roughly 60% a year, resulting in the annual average equity transaction costs of 40 bps. Comparing this number to the back-of-the-envelope housing transaction cost estimates reported above, it seems that even though equity transaction costs are probably somewhat lower, the difference between two asset classes is likely to be small—and no more than 0.5 percentage points a year.
The fact that housing faces much higher costs per each transaction, however, means that the realized housing transaction costs may understate the “shadow” utility cost, which would include the suboptimal allocation choices from staying in the same house and not moving, for example. It might also reduce the volatility of housing returns, making them react more sluggishly to shocks. This means that the relatively modest short-run volatility of housing returns could mask more pronounced fluctuations at lower frequencies. Online Appendix K and Table A.17 compare equity and housing return volatility over longer horizons of up to 20 years. It turns out that the standard deviation of housing returns is always around one-half that of equity returns, regardless of the time horizon, which suggests that housing returns not only have lower short-run volatility but also less pronounced swings at all longer horizons.
2. Leverage
Household-level returns on real estate and equity will be affected by the structure of the household balance sheet, and how these investments are financed. Jordà, Schularick, and Taylor (2016) show that advanced economies in the second half of the twentieth century experienced a boom in mortgage lending and borrowing. This surge in household borrowing not only reflected rising house prices, it also reflected substantially higher household debt levels relative to asset values (and relative to household incomes). The majority of households in advanced economies today hold a leveraged portfolio in their local real estate market. As with any leveraged portfolio, this significantly increases both the risk and the return associated with the investment. And today, unlike in the early twentieth century, houses can be levered much more than equities. The benchmark rent-price ratios from the IPD used to construct estimates of the return to housing refer to rent-price ratios of unleveraged real estate. Consequently, the estimates presented so far constitute only unlevered housing returns of a hypothetical long-only investor, which is symmetric to the way we (and the literature) have treated equities.
However, computing raw returns to housing and equity indices neglects the fact that an equity investment contains embedded leverage. The underlying corporations have balance sheets with both debt and equity liabilities. Thus, reconciliation is needed, and two routes can be taken. For truly comparable raw unlevered returns, equity returns could be delevered. Alternatively, for truly comparable levered returns, housing returns would have to be levered up to factor in the actual leverage (using mortgages) seen on household balance sheets. Is this a big deal in practice? We argue that it does not bias our conclusions significantly based on some elementary calculations.
Consider, for example, the second reconciliation of levering up housing returns. Let the real long-term mortgage borrowing rate be r0, and let α be the leverage of the average house proxied by total mortgages divided by the value of the housing stock. Then we can solve for levered real housing returns TR′ as a function of unlevered real housing returns TR using the formula
. In our data, TR ≃ 7.0% and α ≃ 0.2. Using a long bond return as a proxy for r0 of around 2.5% p.a., this would imply TR′ = 8.1%.38 In other words, for the representative agent the levered housing return is about 110 bps higher than the unlevered housing return (8.1% versus 7%), a small difference. Such adjustments appear to be inconsequential for the main conclusions we present here. In fact, they would bolster one of our central new claims, which is that real housing returns at least match or even exceed real equity returns in the long run when the two are compared on an equal footing.
3. Taxes
When computing equity and housing returns we do not account for taxes. From an investor’s perspective, accounting for taxes is clearly important. Typically, equity capital gains—and, for some countries and periods, even dividend income—have been subject to a capital gains tax. When dividends are not taxed as capital gains, they tend to be taxed as income. In some countries, housing capital gains are subject to capital gains taxes, but owner-occupied houses in particular have been granted exemptions in many cases. Imputed rents of homeowners, unlike dividend income, are almost never taxed. In addition, housing tends to be subject to asset-specific levies in the form of property taxes, documented extensively in Online Appendix Y.
For equities and housing, the level and applicability of taxes has varied over time. For housing, this variation in treatment also extends to assessment rules, valuations, and tax band specifications. As a ballpark estimate, the impact of property taxes would lower real estate returns by less than 1.0 percentage point per year relative to equity (see Online Appendix Y for further details). The various exemptions for homeowners make the impact of capital gains taxes on real estate returns even harder to quantify, but such exemptions also imply that differential tax treatment is unlikely to play an important role in explaining differences in the return between equities and housing.39
Since quantifying the time- and country-varying effect of taxes on returns with precision is beyond the scope of this study, throughout this article we focus on pretax returns from an investor perspective. Importantly, these pretax returns are net of corporate profit taxes, which are netted out before the cash-flow payment to the investor. Studies of returns from an aggregate wealth perspective such as Piketty, Saez, and Zucman (2018) typically compute business equity returns gross of corporate tax. Online Appendix S discusses the impact of adding back corporate taxes on our return data. Equity returns before corporate tax would be roughly 1 percentage point higher than our baseline estimates (Online Appendix Table A.27). This adjustment is, however, very much an upper bound on the housing-equity return differential for the following reasons. First, as noted above, a true like-for-like comparison should also delever equity returns and compare the returns on business and housing wealth. Online Appendix S Table A.27 estimates that first adding back corporate taxes, and then delevering equity returns leaves them approximately equal to the baseline estimates that we report. Second, the total tax burden on the pre–corporate-tax equity returns is likely to be higher than on housing, since in light of the various homeowner exemptions, the post–corporate-tax burden on the two assets appears to be roughly similar. Third, the returns on the two asset classes are similar before 1920, when the corporate tax rate was close to 0.
4. Temporal Aggregation and Return Averaging
The way house and equity price indices are constructed is likely to influence the volatility of the return series. The house price indices used for return calculations (e.g., indices from national statistical agencies) tend to be an average of all transactions in a given year or use a sample of transactions or appraisal values throughout the year. But the equity prices used for return calculations, by the usual convention followed here, compare end-of-year prices of shares. The use of end-of-year rather than yearly average prices mechanically makes equity returns more volatile.
We can assess the magnitude of this effect by constructing an equity return index based on annual averages of daily data, to mimic how housing returns are computed, and then comparing it to a “normal” return using end-of-year index values. For this robustness exercise we use daily MSCI equity returns data for 1970–2015. Table VI presents the end-of-year and yearly average real equity returns in the first two columns, and our yearly average housing returns for the same time period in the third column. Comparing the first two columns shows that yearly averaging lowers the standard deviation of returns by around one-fifth, or 5 percentage points. It also lowers the average return by around 1 percentage point, because the return series is a transformation of the raw price data, and lowering the variance reduces the mean of the return. But the standard deviation of the smoothed yearly average equity series is still almost three times that of housing over the same time period.
Impact of Using End-of-Year versus Yearly Average Asset Prices
| Equity (MSCI index) | Housing (this article) | ||
|---|---|---|---|
| End-of-year | Yearly average | Yearly average | |
| Mean return p.a. | 8.70 | 7.51 | 6.55 |
| Standard deviation | 27.56 | 22.00 | 7.45 |
| Observations | 694 | 694 | 694 |
Notes. Annual global real returns in 16 countries, equally weighted, 1970–2015. End-of-year returns are computed using the return index value for the last day of the year. Yearly average returns are computed using the average index value throughout the year.
Because historical house price data sometimes rely on relatively few transactions, they are likely to be somewhat less smooth than averages of daily data. Therefore Table VI provides an upper bound of the impact of averaging on our return series. Even taking this upper bound at face value, the averaging of house price indices is likely to explain some, but far from all, of the differences in volatility of equity and housing returns.
IV. Risky Rates of Return
At this point the lengthy pilgrimage of Section III ends: the numerous details of how we compiled our data; the many important, but somewhat technical, aspects of data construction; the extensive accuracy checks. In these next sections the focus turns to analysis and interpretation of the data. We examine broad trends and explore their implications for how we think about macroeconomics and finance, confronting the four big themes laid out in the introduction: the long-run behavior of risky returns, safe returns, risk premia, and r minus g.
IV.A. Global Returns
We first turn in Figure VII to a more detailed inspection of the returns on the risky assets, equity and housing. The global returns are GDP-weighted averages of the 16 countries in our sample. Although we do not show the unweighted data, the corresponding figure would look very similar. We smooth the data using decadal moving averages as explained earlier. For example, the observation reported in 1900 is the average of data from 1895 to 1905. Figure VII shows the trends in decadal-average real returns on housing (solid line) and equity (dashed line) for our entire sample. In addition, Figure VIII displays the average rolling decadal correlation of annual risky returns between asset classes, across countries, and with inflation.
Trends in Real Returns on Equity and Housing
Mean returns for 16 countries, weighted by real GDP. Decadal moving averages.
Correlations across Risky Asset Returns
Rolling decadal correlations. The global correlation coefficient is the average of individual countries for the rolling window. Cross-country correlation coefficient is the average of all country pairs for a given asset class. Country coverage differs across time periods.
Risky returns were high in the 1870s and 1880s, fell slowly at first, but then sharply after 1900, with the decade-average real equity returns turning negative during World War I. Risky returns recovered quickly in the 1920s, before experiencing a drop in the the Great Depression, especially for equities. Strikingly, after World War II the trajectories of returns for the two risky asset classes had similar long-run means but over shorter periods diverged markedly from each other.
Equity returns have experienced many pronounced global boom-bust cycles, much more so than housing returns, with average real returns as high as +16% and as low as –4% over entire decades. Equity returns fell in World War II, boomed in the postwar reconstruction, and fell off again in the climate of general macroeconomic instability in the 1970s. Equity returns bounced back following a wave of deregulation and privatization in the 1980s. The next major event was the global financial crisis, which exacted its toll on equities and to some extent housing, as we shall see.
Housing returns, on the other hand, have remained remarkably stable over almost the entire post-World War II period. As a consequence, the correlation between equity and housing returns, depicted in the top left panel of Figure VIII, was highly positive before World War II, but has all but disappeared over the past five decades. The low covariance of equity and housing returns over the long run reveals potential attractive gains from diversification across these two asset classes that economists, up to now, have been unable to measure or analyze.
In terms of relative returns, we see that housing persistently outperformed equity up until World War I, even though both of these asset returns followed a broadly similar temporal pattern. In recent decades, equities have slightly outperformed housing in (arithmetic, not geometric) average but with much higher volatility and cyclicality. Furthermore, upswings in equity prices have generally not coincided with times of low growth or high inflation, when standard asset pricing theory would say high returns would have been particularly valuable.
The top right panel of Figure VIII examines the correlation between risky rates of return and inflation. It shows that equity comoved negatively with inflation in the 1970s, while housing provided a more robust hedge against an unusually rapid surge in consumer prices. In fact, apart from the interwar period, when the world was gripped by a broad deflationary bias, we find that equity returns have comoved negatively with inflation in almost all eras. Moreover, the big downswings in equity returns in the two world wars and the 1970s coincided with periods of generally poor economic performance.
In the past two decades, equity returns have also become highly correlated across countries, as shown by the sharp rise in the degree of cross-country comovement in the bottom left panel of Figure VIII, measured as the average of all country-pair correlations for a given window.40A well-diversified global equity portfolio has thus become less of a hedge against country-specific risk (Quinn and Voth 2008). As is a matter of debate, this may reflect greater freedom to arbitrage and trade across equity markets globally or an increase in the global shocks to which firms, especially those in the typical equity index, are increasingly exposed. In contrast to equities, cross-country housing returns have remained relatively uncorrelated, perhaps because housing assets remain less globally tradable than equities or because they are more exposed to idiosyncratic country-level shocks.
IV.B. Country Returns
Next we explore risky returns in individual countries. Table VII shows returns on equities and housing by country for the full sample and for the post-1950 and post-1980 subsamples. Long-run risky asset returns for most countries are close to 6%–8% a year, a figure we think represents a robust and strong real return to risky capital. Still, the figures also show an important degree of heterogeneity among countries. Many of the countries that experienced large political shocks show lower equity returns. This is the case for Portugal and Spain, which both underwent prolonged civil strife, and France, which undertook a wave of nationalizations in the aftermath of World War II. French equity returns are also negatively affected by the fallout from the world wars and the fallout from an oil crisis in the 1960s (for more detail, see Le Bris and Hautcoeur 2010; Blancheton, Bonin, and Le Bris 2014). In contrast, real equity returns in Finland have been as high as 10% on average throughout the sample. Housing returns also show considerable heterogeneity. Returns on housing have been high on average in the Nordic countries, but low in Italy and Spain. U.S. risky asset returns fall roughly in the middle of the country-specific figures, with equity returns slightly above average and housing returns slightly below. Our estimates of post-World War II U.S. housing returns are in line with those in Favilukis, Ludvigson, and Van Nieuwerburgh (2017).41 The degree of heterogeneity and the relative ranking of returns is broadly similar when comparing the full sample to the post-1950 period.
Real Rates of Return on Equity and Housing
| Full sample | Post-1950 | Post-1980 | ||||
|---|---|---|---|---|---|---|
| Country | Equity | Housing | Equity | Housing | Equity | Housing |
| Australia | 7.79 | 6.37 | 7.53 | 8.29 | 8.70 | 7.16 |
| Belgium | 6.23 | 7.89 | 9.65 | 8.14 | 11.49 | 7.20 |
| Denmark | 7.49 | 8.22 | 9.73 | 7.04 | 13.30 | 5.14 |
| Finland | 10.03 | 9.58 | 12.89 | 11.18 | 16.32 | 9.47 |
| France | 3.21 | 6.39 | 6.01 | 9.68 | 9.61 | 5.78 |
| Germany | 7.11 | 7.82 | 7.53 | 5.30 | 10.07 | 4.13 |
| Italy | 7.25 | 4.77 | 6.09 | 5.55 | 9.45 | 4.57 |
| Japan | 6.00 | 6.54 | 6.21 | 6.74 | 5.62 | 3.58 |
| Netherlands | 6.96 | 7.28 | 9.19 | 8.53 | 11.51 | 6.41 |
| Norway | 5.67 | 8.03 | 7.33 | 9.10 | 12.22 | 9.82 |
| Portugal | 4.51 | 6.31 | 4.84 | 6.01 | 8.60 | 7.15 |
| Spain | 5.83 | 5.21 | 7.75 | 5.83 | 11.96 | 4.62 |
| Sweden | 8.02 | 8.30 | 11.37 | 8.94 | 15.87 | 9.00 |
| Switzerland | 6.51 | 5.63 | 8.37 | 5.64 | 9.29 | 6.19 |
| United Kingdom | 6.83 | 5.44 | 9.10 | 6.57 | 9.11 | 6.81 |
| United States | 8.46 | 6.10 | 8.89 | 5.76 | 9.31 | 5.86 |
| Average, unweighted | 6.67 | 7.26 | 8.30 | 7.47 | 10.78 | 6.43 |
| Average, weighted | 7.12 | 6.72 | 8.19 | 6.40 | 9.08 | 5.50 |
Notes. Average annual real returns. Period coverage differs across countries. Consistent coverage within countries: each country-year observation used to compute the statistics in this table has data for both real housing and equity returns. The average, unweighted and average, weighted figures are, respectively, the unweighted and real-GDP-weighted arithmetic averages of individual country returns.
This country-level evidence reinforces one of our main findings: housing has been as good a long-run investment as equities, possibly better. Housing has offered a similar return to equity in the majority of countries and time periods. In the long run, housing outperformed equities in absolute terms in six countries, and equities outperformed housing in five. Returns on the two assets were about the same in the remaining five countries. After World War II, housing was the best-performing asset class in three countries, and equities were the best-performing in nine.
However, although aggregate total returns on equities exceed those on housing for certain countries and time periods, equities do not outperform housing in simple risk-adjusted terms. Figure IX compares the risk and returns of housing and equities for each country. The left panel plots average annual real returns on housing and equities against their standard deviation. The right panel shows the Sharpe ratios for equities and housing for each country in the sample.42 Housing provides a higher return per unit of risk in each of the 16 countries in our sample, with Sharpe ratios on average more than double those for equities.
Risk and Return of Equity and Housing
Left: average real return p.a. and standard deviation. Right: Sharpe ratios, measured as
mean return and σi standard deviation. Sixteen countries. Consistent coverage within each country.
IV.C. Decomposition of Returns
To further look into the underlying drivers of housing and equity returns, we decompose them into the capital gain (price) and yield (dividend or rent) components. To be consistent with the data in Section III and Table II, we decompose real total return into real capital gain—that is, the price change net of inflation—and dividend or rental yield—that is, the nominal yield as proportion of the previous year’s share or house price.43 Yet caveats arise. In principle, it is not entirely clear whether inflation should be subtracted from the capital gain or yield component. Moreover, firms may buy back shares to generate low-tax capital gains instead of paying out higher-taxed dividends; thus, the manner of distribution of total returns may not be invariant to circumstances.
Table VIII decomposes equity and housing returns into capital gains and dividends or rents, for the full cross-country sample and the period after 1950. Over the full sample, most of the real return is attributable to the yield. Dividends account for roughly 60% of real equity returns, and rents for roughly 80% of real housing returns. In terms of geometric means (Table VIII, third row), almost all of both equity and housing returns are attributable to, respectively, dividend and rental income. After 1950, capital gains become more important for both equities and housing. For equities, real capital gains account for the majority of the total return after 1950, and for housing for roughly one-third.
Total Return Components for Equity and Housing
| Equity | Housing | |||||
|---|---|---|---|---|---|---|
| Real | Real | Real | Real | |||
| capital | Dividend | total | capital | Rental | total | |
| gain | income | return | gain | income | return | |
| Full sample | ||||||
| Mean return p.a. | 2.78 | 4.17 | 6.82 | 1.61 | 5.50 | 6.92 |
| Standard deviation | 21.37 | 1.74 | 21.89 | 9.87 | 2.05 | 10.40 |
| Geometric mean | 0.57 | 4.16 | 4.58 | 1.15 | 5.48 | 6.43 |
| Observations | 1,707 | 1,707 | 1,707 | 1,707 | 1,707 | 1,707 |
| Post-1950 | ||||||
| Mean return p.a. | 4.73 | 3.80 | 8.36 | 2.39 | 5.22 | 7.38 |
| Standard deviation | 23.70 | 1.81 | 24.24 | 8.59 | 1.93 | 8.95 |
| Geometric mean | 2.03 | 3.79 | 5.62 | 2.06 | 5.21 | 7.04 |
| Observations | 995 | 995 | 995 | 995 | 995 | 995 |
Notes. Average annual real capital gain, dividend, or rental income, and total return across 16 countries, unweighted. Period coverage differs across countries. Consistent coverage within countries: each country-year observation used to compute the statistics in this table has data for both equity and housing returns, capital gains, and yields. Dividend and rental income are expressed in percent of the previous year’s asset price.
The importance of dividends and rents is partly a matter of convention. Online Appendix N and Online Appendix Table A.20 computes the equivalent decomposition for nominal returns, and finds that the capital gain versus dividend/rental income split is then closer to roughly 50/50. Nevertheless, without dividends or rents, the real returns on both assets would be low, especially in geometric mean terms. This is consistent with the existing literature on real house prices: Shiller (2000) documents that house prices in the United States moved in line with inflation before the 2000s bubble, and Knoll, Schularick, and Steger (2017) show that real house prices in advanced economies were more or less flat before 1950. This is also true in our data: the pre-1950 annual real housing capital gains are just 0.5%. Post-1950 capital gains are somewhat higher at 2.5%, but still only half the magnitude of the rental yields. Adding rents to the equation radically changes the picture and brings the long-run housing total return close to 7%. Interestingly, the broad picture is similar for equities: the real equity capital gain before 1950 is, on average, just 0.4%, compared with 4.7% a year after 1950. However, the contribution of dividend and rental income means that housing and equity returns were high both before and after 1950.
Although most of the return can be attributed to dividends and rents, almost all of the volatility comes from equity and house prices, that is, the capital gains component. Table VIII, second row shows that both dividends and rents are very stable, with a standard deviation of dividend/rental yields of about 2%. Prices, on the contrary, move around much more, with a standard deviation roughly equal to that of total returns (21.4% for equities and 9.9% for housing). The higher variability of equity returns and the superior risk-adjusted performance of housing can therefore largely be attributed to the lower volatility of house prices compared to those of equities. An additional factor is that capital gains—the more volatile component—account for a relatively larger share of equity returns.
Because aggregate equity prices are subject to large and prolonged swings, a representative investor would have to hold on to his equity portfolio for longer to ensure a high real return. Aggregate housing returns, on the other hand, are more stable because swings in aggregate house prices are generally less pronounced. National aggregate housing portfolios have had comparable real returns to national aggregate equity portfolios, but with only half the volatility.
Table IX examines the relative importance of capital gains versus dividends/rents at the country level (figures are arithmetic means and standard deviations). Additionally we report the share of real total return accounted for by capital gains. The fact that the majority of housing returns come from yields is true across all countries. The lowest real capital gains are observed in Switzerland, and the highest in Finland, but none exceed 3% a year in the full sample. Rents are relatively more important in the United States, accounting for roughly 90% of returns, but this is not unusual: Denmark, Sweden, and Switzerland have the same share of capital gains as the United States. For equities, the picture is more mixed. Seven countries, including the United States, have a roughly 50/50 split between real capital gain and dividend yield shares. Other countries record low or negative real capital gains over the full sample, especially in geometric mean terms (see Online Appendix Table A.22).
Total Return Components for Equity and Housing by Country
| Equity | Housing | |||||||
|---|---|---|---|---|---|---|---|---|
| Real | Real | Capital | Real | Real | Capital | |||
| capital | Dividend | total | gain | capital | Rental | total | gain | |
| gain | income | return | share | gain | income | return | share | |
| Australia | 3.06 | 4.90 | 7.79 | 0.38 | 2.53 | 3.99 | 6.37 | 0.24 |
| (16.30) | (1.08) | (16.94) | (11.94) | (0.92) | (11.92) | |||
| Belgium | 2.53 | 3.83 | 6.23 | 0.40 | 1.95 | 6.15 | 7.89 | 0.14 |
| (22.92) | (1.64) | (23.61) | (15.05) | (1.46) | (15.51) | |||
| Denmark | 2.71 | 4.95 | 7.49 | 0.35 | 1.26 | 7.13 | 8.22 | 0.08 |
| (16.14) | (2.09) | (16.45) | (7.02) | (2.42) | (7.60) | |||
| Finland | 5.19 | 5.08 | 10.03 | 0.51 | 2.82 | 7.14 | 9.58 | 0.17 |
| (31.18) | (1.95) | (31.93) | (14.61) | (2.86) | (15.62) | |||
| France | −0.37 | 3.73 | 3.21 | 0.09 | 1.55 | 5.09 | 6.39 | 0.13 |
| (21.57) | (1.33) | (22.14) | (9.39) | (1.14) | (10.03) | |||
| Germany | 2.74 | 4.08 | 7.11 | 0.40 | 1.86 | 6.03 | 7.82 | 0.13 |
| (20.99) | (1.58) | (21.72) | (9.24) | (2.61) | (10.16) | |||
| Italy | 3.78 | 3.61 | 7.25 | 0.51 | 1.45 | 3.49 | 4.77 | 0.18 |
| (27.99) | (1.34) | (28.42) | (9.28) | (1.59) | (9.61) | |||
| Japan | 3.12 | 2.65 | 6.00 | 0.54 | 2.00 | 4.70 | 6.54 | 0.18 |
| (18.94) | (1.77) | (19.15) | (7.99) | (1.24) | (8.41) | |||
| Netherlands | 3.38 | 4.87 | 8.10 | 0.41 | 1.75 | 5.96 | 7.51 | 0.13 |
| (19.21) | (1.57) | (19.61) | (8.22) | (1.68) | (8.76) | |||
| Norway | 1.61 | 4.21 | 5.67 | 0.28 | 1.49 | 6.72 | 8.03 | 0.10 |
| (19.33) | (1.60) | (19.82) | (8.26) | (1.19) | (8.70) | |||
| Portugal | 2.92 | 2.28 | 5.11 | 0.56 | 1.13 | 4.47 | 5.21 | 0.12 |
| (34.34) | (1.22) | (34.73) | (9.26) | (1.98) | (9.37) | |||
| Spain | 1.80 | 4.53 | 5.83 | 0.28 | 1.26 | 4.16 | 5.21 | 0.13 |
| (20.48) | (2.30) | (21.15) | (11.59) | (1.60) | (12.00) | |||
| Sweden | 4.08 | 4.12 | 8.02 | 0.50 | 1.39 | 7.12 | 8.30 | 0.09 |
| (19.54) | (1.05) | (20.03) | (8.46) | (1.61) | (8.88) | |||
| Switzerland | 3.17 | 3.20 | 6.27 | 0.50 | 0.81 | 4.54 | 5.24 | 0.08 |
| (20.61) | (1.46) | (20.73) | (6.50) | (0.62) | (6.74) | |||
| United Kingdom | 2.48 | 4.53 | 6.83 | 0.35 | 1.63 | 3.94 | 5.44 | 0.17 |
| (19.12) | (1.39) | (19.73) | (8.94) | (0.86) | (9.15) | |||
| United States | 4.19 | 4.38 | 8.46 | 0.49 | 0.90 | 5.33 | 6.10 | 0.08 |
| (18.90) | (1.57) | (19.17) | (7.84) | (0.75) | (8.12) | |||
Notes. Arithmetic average of annual real capital gain, dividend or rental income, and total return, full sample. Standard deviation in parentheses. Period coverage differs across countries. Consistent coverage within countries: each country-year observation used to compute the statistics in this table has data for both equity and housing returns, capital gains, and yields. Dividend and rental income are expressed as a percentage of the previous year’s asset price. Capital gain share is the proportion of real total return attributable to real capital gains.
V. Safe Rates of Return
Turning to safe asset returns, Figure X shows the trends in real returns on government bonds (solid line) and bills (dashed line) since 1870. Again, returns are GDP-weighted averages of the 16 countries in our sample; the corresponding unweighted figure would look very similar. We smooth the data using decadal moving averages, as explained earlier.
Trends in Real Returns on Bonds and Bills
Mean returns for 16 countries, weighted by real GDP. Decadal moving averages.
Three striking features of Figure X deserve comment. First, low real rates and, in fact, negative real rates have been relatively common in modern financial history. Second, for the most part, returns to long-term and short-term safe assets have tracked each other very closely—with a premium of about 1% that has widened considerably since the well-documented decline of the mid-1980s (see, e.g., Holston, Laubach, and Williams 2017). Third, a major stylized fact leaps out once we compare the safe rates of return in Figure X to the risky rates of return in Figure VII. Prior to World War II, real returns on housing, safe assets, and equities followed remarkably similar trajectories. After World War II this was no longer the case.
Safe rates are far from stable in the medium term. There is enormous time-series, as well as cross-country, variability. In fact, real safe rates appear to be as volatile as real risky rates (sometimes more volatile), a topic we return to in the next section. Considerable variation in the risk premium often comes from sharp changes in safe real returns, not from real returns on risky assets.
Two four-decade-long declines in real rates stand out: (i) from 1870 to World War I (with a subsequent further collapse during the war), and (ii) the well-documented decline that started in the mid-1980s. We could add to this list the briefer, but more dramatic decline that followed the Great Depression into World War II. Some observers have therefore interpreted the recent downward trend in safe rates as a sign of a new era of “secular stagnation” (see, e.g., Summers 2014).
However, in contrast to 1870–1913 and the 1930s, the more recent decline is characterized by a much higher term premium—a feature with few precedents in our sample.44 There are other periods in which real rates remained low, such as in the 1960s. They were pushed below zero, particularly for the longer-tenor bonds, during the 1970s inflation spike, although here too term premiums remained relatively tight. Returns also dipped dramatically during both world wars. This is perhaps to be expected: demand for safe assets spikes during disasters, although the dip may also reflect periods of financial repression and high inflation that usually emerge during times of conflict and often persist into peacetime. Thus, from a broad historical perspective, high rates of return on safe assets and high term premiums are more the exception than the rule.
Summing up, over more than 140 years from the late nineteenth to the twenty-first century, real returns on safe assets have been low—on average 1% for bills and 2.5% for bonds—relative to alternative investments. Although the return volatility—measured as annual standard deviation—is lower than that of housing and equities, these assets offered little protection during high-inflation eras and during the two world wars, both periods of low consumption growth.
Again moving on to examine correlations, Figure XI explores additional key moments of the data. The top left panel plots the correlation between real bond and real bill returns, again using decadal rolling windows and computed as the cross-sectional average of correlations. In parallel to our discussion of the term premium, real returns on bonds and bills have been highly correlated for most of the sample up until the 1960s. From the 1970s onward, the era of fiat money and higher average inflation, this correlation has become much weaker, and near zero at times, coinciding with a widening term premium
Correlations across Safe Asset Returns
Color version available online. Rolling decadal correlations. The global correlation coefficient is the average of individual countries for the rolling window. The cross-country correlation coefficient is the average of all country pairs for a given asset class. Country coverage differs across time periods.
The top right panel of Figure XI displays the correlation between nominal safe asset returns and inflation for real bond and real bill returns. The figure shows that safe assets provided more of an inflation hedge starting in the 1970s, around the start of the era of modern central banking. However, as Figure X showed, bonds and bills have experienced prolonged periods of negative real returns—both during wartime inflations and in the high-inflation period of the late 1970s. Although safe asset rates usually comove positively with inflation, they do not always compensate the investor fully.
The bottom panel of Figure XI displays the cross-correlation of safe returns over rolling decadal windows, averaged for all country-pair combinations, to examine how much risk can be diversified with debt instruments. Cross-country real safe returns have exhibited positive comovement throughout history. The degree of comovement shows a few marked increases in World War I and in the 1930s. The effect of these major global shocks on individual countries seems to have resulted in a higher correlation of cross-country asset returns.
Turning to cross-sectional features, Table X shows country-specific safe asset returns for three samples: all years, post-1950, and post-1980. Here the experiences of a few countries stand out. In France, real bill returns have been negative when averaged over the full sample. In Portugal and Spain, they have been approximately zero. In Norway, the average return on bills has been negative for the post-1950 sample. However, most other countries have experienced reasonably similar returns on safe assets, in the ballpark of 1%–3%.
Real Rates of Return on Bonds and Bills
| Full Sample | Post-1950 | Post-1980 | ||||
|---|---|---|---|---|---|---|
| Country | Bills | Bonds | Bills | Bonds | Bills | Bonds |
| Australia | 1.29 | 2.24 | 1.32 | 2.45 | 3.23 | 5.85 |
| Belgium | 1.21 | 3.01 | 1.61 | 3.86 | 2.51 | 6.24 |
| Denmark | 3.08 | 3.58 | 2.18 | 3.50 | 2.80 | 7.13 |
| Finland | 0.64 | 3.22 | 0.63 | 4.86 | 2.61 | 5.76 |
| France | −0.47 | 1.54 | 0.96 | 2.97 | 2.24 | 6.96 |
| Germany | 1.51 | 3.15 | 1.86 | 3.70 | 1.97 | 4.23 |
| Italy | 1.20 | 2.53 | 1.30 | 2.83 | 2.42 | 5.85 |
| Japan | 0.68 | 2.54 | 1.36 | 2.83 | 1.48 | 4.53 |
| Netherlands | 1.37 | 2.71 | 1.04 | 2.14 | 2.08 | 5.59 |
| Norway | 1.10 | 2.55 | −0.26 | 1.94 | 1.50 | 5.62 |
| Portugal | −0.01 | 2.23 | −0.65 | 1.59 | 0.65 | 6.25 |
| Spain | −0.04 | 1.41 | −0.32 | 1.21 | 2.20 | 5.72 |
| Sweden | 1.77 | 3.25 | 0.82 | 2.71 | 1.52 | 6.60 |
| Switzerland | 0.89 | 2.41 | 0.12 | 2.33 | 0.33 | 3.35 |
| United Kingdom | 1.16 | 2.29 | 1.14 | 2.63 | 2.70 | 6.67 |
| United States | 2.23 | 2.85 | 1.43 | 2.77 | 1.91 | 5.90 |
| Average, unweighted | 1.14 | 2.61 | 0.91 | 2.77 | 2.01 | 5.77 |
| Average, weighted | 1.34 | 2.51 | 1.23 | 2.70 | 1.98 | 5.64 |
Notes. Average annual real returns. Period coverage differs across countries. Consistent coverage within countries: each country-year observation used to compute the statistics in this table has data for both real bill and bond returns. The average, unweighted and average, weighted figures are, respectively, the unweighted and real-GDP-weighted arithmetic averages of individual country returns.
Aside from the investor perspective discussed already, safe rates of return have important implications for government finances, as they measure the cost of raising and servicing government debt. What matters for this is not the level of real return per se, but its comparison to real GDP growth, or rsafe − g. If the rate of return exceeds real GDP growth, rsafe > g, reducing the debt/GDP ratio requires continuous budget surpluses. When rsafe is less than g, however, a reduction in debt/GDP is possible even with the government running modest deficits. Existing evidence points to rsafe < g being the norm rather than the exception, both in recent years and broader historical data (Ball, Elmendorf, and Mankiw 1998; Mehrotra 2017).
Figure XII plots the representative “safe rate of return” as the arithmetic average of bond and bill returns (dashed line) alongside real GDP growth (solid line), again as decadal moving averages. Starting in the late nineteenth century, safe rates were higher than GDP growth, meaning that any government wishing to reduce debt had to run persistent budget surpluses. Indeed, this was the strategy adopted by Britain to pay off the debt incurred during the Napoleonic War (Crafts 2016). The two world wars saw low real returns, but nevertheless a large debt accumulation to finance the wartime effort. The aftermath of these two wars, however, offered vastly different experiences for public finances. After World War I, safe returns were high and growth low, requiring significant budgetary efforts to repay the war debts. This was particularly difficult for many countries given the large interlocking reparations imposed by the Treaty of Versailles, and the turbulent macroeconomic environment at the time. After World War II, on the contrary, high growth and inflation helped greatly reduce the value of national debt, creating rsafe − g gaps as large as –10 percentage points.
Trends in the Real Return on Safe Assets and GDP Growth
Mean returns and GDP growth for 16 countries, weighted by real GDP. Decadal moving averages. The safe rate of return is an arithmetic average of bonds and bills.
More recently, the Great Moderation saw a reduction in inflation rates and a corresponding increase in the debt financing burden, whereas the impact of rsafe − g in the aftermath of the Global Financial Crisis remains broadly neutral, with the two rates roughly equal. On average throughout our sample, the real growth rate has been around 1 percentage point higher than the safe rate of return (3% growth versus 2% safe rate), meaning that governments could run small deficits without increasing the public debt burden.
In sum, real returns on safe assets have been quite low across the advanced countries over the past 150 years. In fact, for some countries, these returns have often been persistently negative. Periods of unexpected inflation, in war and peace, have often diluted returns, and flights to safety may have depressed safe returns even further in the more turbulent periods of global financial history. The low return for investors has, on the flipside, implied a low financing cost for governments, which was particularly important in reducing the debts incurred during World War II.
VI. Risky versus Safe Returns
Having established the general trends in each risky and safe asset class, we turn to examine broader patterns of returns across the different asset classes. We start by comparing returns on risky and safe assets. Figure XIII depicts the trends in global safe and risky asset returns, again using decadal moving averages of GDP-weighted global return series.
Global Real Risky versus Real Safe Return
Mean returns for 16 countries, weighted by real GDP. Decadal moving averages. Within each country, the real risky return is a weighted average of equities and housing, and the safe return is a weighted average of bonds and bills. The within-country weights correspond to the shares of the respective asset in the country’s wealth portfolio. Risk premium = risky return – safe return.
The risky return in each country is a weighted average of housing and equity returns, with weights corresponding to equity market capitalization and housing wealth in each respective country. The safe return is a simple unweighted average of bonds and bills.45 The left panel of Figure XIII shows the risky and safe asset returns, and the right panel depicts the risk premium, calculated as the risky minus safe difference.
As in Sections IV and V, the data presented here measure ex post returns and risk premiums, inclusive of capital gains. For some of the debates we touch on, however, a forward-looking expected return measure would have been preferable. Realized returns are likely to fall below ex ante return expectations during periods of large negative shocks, such as the two world wars, and rise above them in times of high capital gains, such as that between 1980 and today. Long-run data on expected returns are, however, difficult to obtain. Our focus on long-run trends, to an extent, allows us to look through some of the unexpected shocks that drive a wedge between ex ante and ex post returns. Nevertheless, the discussion in this section should be treated with a note of caution.
Both risky and safe returns were high during the nineteenth century but had been gradually declining in the run up to World War I, after which they declined sharply, as is to be expected. After the war, returns were recovering during the 1920s. From 1930 onward, the risky return stayed high and relatively stable, whereas the safe return dropped sharply and remained low until the late 1970s, before increasing and falling back again over the past three decades. These findings have implications for current debates around secular stagnation and the pricing, or mispricing, of risk.
Secular stagnation is associated with low rates of return, driven by an excess of savings or a general unwillingness to borrow and invest. These in turn reflect a variety of potential factors, including (i) lower rates of productivity growth; (ii) lower fertility and mortality rates; (iii) a decline in the relative price of investment goods; (iv) greater firm-level market power; and (v) higher income inequality (Rachel and Smith 2015; Eggertsson, Mehrotra, and Robbins 2017; Thwaites 2015).
Indeed, we can see that the safe return fell sharply during the 1930s, when Hansen (1939) originally proposed the secular stagnation hypothesis. That time also coincided with a demographic bust and was preceded by a big rise in income inequality in the run-up to the Great Depression. The safe return has been falling again since the mid-1980s, as many have noted.46 Understandably, this has led some observers to suggest that advanced economies are again in danger of entering secular stagnation, for example, Summers (2014) and Eggertsson and Mehrotra (2014).
The picture changes radically when we consider the trend in risky returns in addition to safe returns. Unlike safe returns, risky returns have remained high and broadly stable through the best part of the past 100 years, and show little sign of a secular decline. Turning back to the trend for safe assets, even though the safe return has declined recently, much as it did at the start of our sample, it remains close to its historical average. These two observations call into question whether secular stagnation is quite with us. The high and stable risky return coupled with falling safe rates could also be consistent with the notion of a “safety trap” brought about by the relative shortage of safe assets (Caballero and Farhi 2017). However, with risk premiums still not far off their historical averages, the evidence for a safety trap is thus far also not clear-cut.
We turn to examine the long-run developments in the ex post risk premium, that is, the spread between safe and risky returns (right panel of Figure XIII). This spread was low and stable at around 5 percentage points before World War I. It rose slightly after World War I, before falling to an all-time low of near zero by around 1930. The decades following the onset of World War II saw a dramatic widening in the risk premium, with the spread reaching its historical high of around 14 percentage points in the 1950s, before falling back to around its historical average.
Interestingly, the period of high risk premiums coincided with an era of few systemic banking crises. In fact, not a single such crisis occurred in our advanced-economy sample between 1946 and 1973. By contrast, banking crises appear to have been relatively more frequent when risk premiums were low. This finding speaks to the recent literature on the mispricing of risk around financial crises. Among others, Krishnamurthy and Muir (2017) argue that when risk is underpriced, that is, risk premiums are excessively low, severe financial crises become more likely.
The long-run trends in risk premiums presented here seem to confirm this hypothesis. Online Appendix F further examines how these long-run movements in the risk premium, and the returns on the individual risky and safe asset classes, correspond to the changing monetary regimes, and finds, in accordance with Figure XIII, that the risk premium during the Bretton Woods fixed exchange rate era was unusually high by historical standards, driven largely by the low, even negative, safe asset returns but also by reasonably high housing returns.
Table XI zooms in to examine the evolution of safe and risky asset returns across different countries, as well as time periods. To enable a comparison with the aggregate trends in Figure XIII, we split the post-World War II period into two subperiods: 1950–1980, when global risk premiums were high and global safe returns low, and post-1980, which saw an initial recovery, and subsequent decline in global safe returns.
Real Risky and Safe Asset Returns across Countries and Time
| Full sample | 1950–1980 | Post-1980 | ||||
|---|---|---|---|---|---|---|
| Risky | Safe | Risky | Safe | Risky | Safe | |
| Country | return | return | return | return | return | return |
| Australia | 6.96 | 1.77 | 6.51 | −1.34 | 7.71 | 4.54 |
| Belgium | 8.31 | 1.82 | 9.68 | 1.05 | 7.99 | 4.38 |
| Denmark | 8.02 | 3.05 | 8.57 | 0.49 | 6.84 | 4.97 |
| Finland | 10.87 | 2.16 | 13.47 | 1.28 | 13.06 | 4.18 |
| France | 6.54 | 0.54 | 12.33 | −1.15 | 6.61 | 4.60 |
| Germany | 7.90 | 3.34 | 7.00 | 1.77 | 5.20 | 3.10 |
| Italy | 5.32 | 2.28 | 7.08 | −0.83 | 5.21 | 4.14 |
| Japan | 6.79 | 1.29 | 10.86 | 0.05 | 4.81 | 3.00 |
| Netherlands | 7.30 | 1.31 | 10.26 | −0.89 | 7.42 | 3.83 |
| Norway | 7.96 | 1.59 | 7.75 | −2.34 | 10.65 | 3.56 |
| Portugal | 6.46 | 0.45 | 5.19 | −3.30 | 7.41 | 3.45 |
| Spain | 5.39 | 0.68 | 7.27 | −3.56 | 5.46 | 3.96 |
| Sweden | 8.52 | 2.35 | 8.67 | −1.12 | 11.42 | 4.06 |
| Switzerland | 6.51 | 1.57 | 6.07 | 0.25 | 7.76 | 1.84 |
| United Kingdom | 6.35 | 1.51 | 8.33 | −1.36 | 7.66 | 4.69 |
| United States | 7.12 | 1.92 | 6.44 | −0.32 | 7.28 | 3.91 |
| Average, unweighted | 7.44 | 1.88 | 8.48 | −0.81 | 7.65 | 3.89 |
| Average, weighted | 7.22 | 1.89 | 7.88 | −0.56 | 6.66 | 3.81 |
Notes. Average annual real returns. Real risky return is a weighted average of equity and housing, and safe return is a weighted average of bonds and bills. The weights correspond to the shares of the respective asset in the country’s wealth portfolio. Period coverage differs across countries. Coverage is consistent within countries: each country-year observation used to compute the statistics in this table has data for both the risky and safe return. The average, unweighted and average, weighted figures are, respectively, the unweighted and real-GDP-weighted arithmetic averages of individual country returns.
The vast majority of countries in our sample follow similar patterns. The risky return is largely stable across time, even though it varies somewhat across countries: from just over 5% in Italy and Spain to 11% in Finland. Risk premiums were at or near their highest level in almost every country during the period 1950–1980, largely due to low returns on safe assets. The real safe rate of return was close to zero or negative for the majority of the countries in the sample, with the lowest level of –3.5% observed in Spain and Portugal, and only Belgium, Finland, and Germany experiencing robustly positive real returns. Meanwhile, risky returns were also somewhat above their long-run level in a number of countries, but the differences are relatively smaller than those for safe rates. The post-1980 period saw a recovery in safe returns across the board, with the recent downward trend not yet apparent in these longer-run period averages. Risky returns, meanwhile, were close to their historical levels in most countries, with only Japan experiencing a strong decline following the bursting of its asset price bubble in the 1990s.
We now turn to examine the correlations between risky and safe returns, which are displayed in Figure XIV. The top left panel of this figure shows the rolling decadal correlation between the risky and safe returns, calculated as the average of rolling correlations in individual countries in a similar fashion to the calculations in Figure VIII. Throughout most of the historical period under consideration, risky and safe returns had been positively correlated. In other words, safe assets have not generally provided a hedge against risk because safe returns were low when risky returns were low—in particular during both world wars—and vice versa. This positive correlation has weakened over the more recent decades and turned negative from the 1990s onward. This suggests that safe assets have acted as a better hedge for risk during both the Great Moderation and the recent Global Financial Crisis.
Correlations across Risky and Safe Asset Returns
Rolling decadal correlations. The global correlation coefficient is the average of individual countries for the rolling window. Cross-country correlation coefficient is the average of all country pairs for a given asset class. Country coverage differs across time periods.
The top right panel of Figure XIV shows the comovement of risky and safe nominal returns with inflation. Mirroring our findings presented in the preceding sections, safe returns have tended to comove more strongly with inflation, particularly during the post-World War II period. Moving to cross-country correlations depicted in the bottom two panels of Figure XIV, historically safe returns in different countries have been more correlated than risky returns. This has reversed over the past decades as cross-country risky returns have become substantially more correlated. This seems to be mainly driven by a remarkable rise in the cross-country correlations in risk premiums, depicted in the bottom right panel of Figure XIV. This increase in global risk comovement may pose new challenges to the risk-bearing capacity of the global financial system, a trend consistent with other macro indicators of risk-sharing (Jordà, Schularick, and Taylor 2017).
VII. r versus g
Our analysis provides insights into the debate on inequality. Piketty (2014) and Piketty and Zucman (2014) argue that inequality and wealth-to-income ratios in advanced economies have followed a U-shaped pattern over the past century and a half. They further hypothesize that wealth inequality may continue to rise in the future, along with a predicted decline in the rate of economic growth. The main theoretical argument for this comes about from a simple relation: r > g. In their approach, a higher spread between the real rate of return on wealth, denoted r, and the rate of real GDP growth, g, tends to magnify the steady-state level of wealth inequality. Benhabib and Bisin (2016) show that in a wide class of models featuring stochastic returns to wealth, a higher gap between r and g increases the Pareto index of the steady-state wealth distribution, making it more unequal.
Of course, this is not the only channel through which rates of return can impact the wealth distribution. Rate of return differentials between asset classes can affect the wealth distribution if there are systematic differences in the portfolio composition between rich and poor households, as Kuhn, Schularick, and Steins (2017) show, or if rates of returns vary with portfolio size, as stressed by Piketty (2014). Studying administrative Swedish data, Bach, Calvet, and Sodini (2016) find that wealthy households earn higher returns on their portfolios, and Fagereng et al. (2016) use Norwegian tax data to document substantial heterogeneity in wealth returns. Rates of return on wealth are beginning to receive attention in the theoretical literature. For instance, Benhabib and Bisin (2016) point to return differences of assets as a potential channel to explain diverging trends between income and wealth inequality, and Garbinti, Goupille-Lebret, and Piketty (2017) show that asset price effects played an important role in shaping the French wealth distribution over the past 200 years. Furthermore, wealth inequality may depend not only on the magnitude of r in relation to g but also on return volatility. Higher return volatility can increase the dispersion of wealth outcomes and make the distribution of wealth more unequal.
To bring our data to bear on these debates, we construct a measure of the world’s real return on wealth as a weighted average of real returns on bonds, equities, and housing—reflecting the typical portfolio of a private household end-investor. We then compare this measure to the rate of real GDP growth of economies over the long run. Importantly, our approach differs from Piketty (2014) in that we rely on annual returns from observed market prices and yields for each individual asset class, rather than implicit returns derived from aggregate balance sheet data at selected benchmark dates. This, we think, is more robust and provides a vital cross check for the core argument.
Similarly to the risky returns in Section VI, we weight the individual returns by the size of the respective asset portfolio: stock market capitalization, housing wealth, and public debt (divided equally between bonds and bills).47Figure XV displays the long-run trends in the global real rate of return on wealth (dashed line) and the global real GDP growth rate (solid line) since the late nineteenth century, again using decadal moving averages of GDP-weighted data.
The Real Return on Wealth and Real GDP Growth
Mean returns and real GDP growth for 16 countries, weighted by real GDP. Decadal moving averages. Within each country, the real return on wealth is a weighted average of bonds, bills, equity, and housing. The within-country weights correspond to the shares of the respective asset in each country’s wealth portfolio.
Our data show that the trend long-run real rate of return on wealth has consistently been much higher than the real GDP growth rate. Over the past 150 years, the real return on wealth has substantially exceeded real GDP growth in 13 decades and has only been below GDP growth in the 2 decades corresponding to the two world wars. That is, in peacetime, r has always exceeded g. The gap between r and g has been persistently large. Since 1870, the weighted average return on wealth (r) has been about 6.0%, compared to a weighted average real GDP growth rate (g) of 3.0%, with the average r − g gap of 3.0 percentage points, which is about the same magnitude as the real GDP growth rate itself. The peacetime gap between r and g has been larger still, averaging around 3.8 percentage points.
Table XII shows how the rate of return on wealth and the GDP growth rate have varied across different countries and time periods. Despite some variation, the positive gap between r and g is a persistent feature of the data: r is bigger than g in every country and every time period that we consider. The last few decades prior to the Global Financial Crisis saw a general widening of this gap, mirroring the aggregate pattern shown in Figure XV.
The Return on Wealth and GDP Growth Across Countries and Time
| Full sample | Post-1950 | Post-1980 | ||||
|---|---|---|---|---|---|---|
| Return on | GDP | Return on | GDP | Return on | GDP | |
| Country | wealth | growth | wealth | growth | wealth | growth |
| Australia | 5.91 | 3.51 | 7.39 | 3.73 | 7.53 | 3.19 |
| Belgium | 6.38 | 2.32 | 7.29 | 2.68 | 6.90 | 2.17 |
| Denmark | 7.37 | 2.70 | 7.21 | 2.51 | 6.62 | 1.60 |
| Finland | 9.76 | 3.49 | 11.92 | 3.16 | 11.81 | 2.16 |
| France | 4.92 | 2.55 | 7.76 | 3.17 | 6.29 | 1.92 |
| Germany | 7.07 | 2.81 | 5.26 | 2.80 | 4.72 | 2.40 |
| Italy | 5.08 | 3.82 | 5.07 | 3.30 | 5.01 | 1.37 |
| Japan | 5.59 | 4.18 | 6.35 | 4.20 | 4.23 | 2.09 |
| Netherlands | 5.33 | 3.16 | 6.67 | 3.21 | 6.71 | 2.29 |
| Norway | 6.86 | 3.06 | 7.67 | 3.45 | 9.35 | 2.80 |
| Portugal | 5.87 | 3.39 | 5.65 | 3.48 | 6.99 | 2.13 |
| Spain | 4.58 | 3.21 | 5.50 | 4.03 | 5.34 | 2.56 |
| Sweden | 7.41 | 2.89 | 8.69 | 2.86 | 9.87 | 2.36 |
| Switzerland | 5.63 | 2.33 | 5.98 | 2.69 | 7.03 | 1.95 |
| United Kingdom | 4.75 | 2.09 | 5.90 | 2.49 | 7.23 | 2.45 |
| United States | 6.03 | 3.38 | 5.91 | 3.33 | 6.58 | 2.82 |
| Average, unweighted | 6.30 | 2.86 | 6.92 | 3.23 | 7.01 | 2.26 |
| Average, weighted | 5.98 | 3.04 | 6.09 | 3.33 | 6.08 | 2.48 |
Notes. Average annual real returns. Real return on wealth is a weighted average of bonds, bills, equity, and housing. The weights correspond to the shares of the respective asset in each country’s wealth portfolio. Period coverage differs across countries. Coverage is consistent within countries: each country-year observation used to compute the statistics in this table has data for both the real return on wealth and the real GDP growth rate. The average, unweighted and average, weighted figures are, respectively, the unweighted and real-GDP-weighted arithmetic averages of individual country returns.
As previously discussed, returns on housing play an important part in this story—but with scant data until now, their exact role was unclear. The high level of housing returns that we have uncovered serves to push up the level of r, and thus, potentially, wealth inequality. But what is the counterfactual? We need to bear in mind that housing wealth is more equally distributed than, for instance, equities (see, e.g., Kuhn, Schularick, and Steins 2017), and returns on housing are less volatile than those on shares—with both factors serving to flatten the distribution of wealth changes, making the overall impact of housing returns on wealth inequality less clear-cut and offering substantial scope for further research.
Rognlie (2015) notes that recent trends in wealth and income could be influenced primarily by what has happened in housing. Real house prices have experienced a dramatic increase in the past 40 years, coinciding with the rapid expansion of mortgage lending (Knoll, Schularick, and Steger 2017; Jordà, Schularick, and Taylor 2015, 2016). This is very much evident from Table IX. Measured as a ratio to GDP, rental income has been growing, as Rognlie (2015) argues. However, the rental yield has declined slightly—given the substantial increase in house prices—so that total returns on housing have remained pretty stable, as we have discussed. In this sense, recent housing trends have diverged little.
Our data allow us to more formally examine whether movements in the r − g gap are more closely related to return fluctuations or movements in the real GDP growth rate. Online Appendix Q and Table A.26 document the correlations between r − g and g, and r − g and r across different time horizons, for the full sample and the period after 1950. Overall, the correlation between r − g and g is negative, and somewhat stronger at longer horizons, with the correlation coefficients ranging between –0.2 and –0.6 depending on the historical subperiod and time window. At the same time, the r − g gap is even more robustly related to changes in the return on wealth r, with a positive correlation between the two and a correlation coefficient of around 0.9, both over the short and long run. This suggests that both falling GDP growth and higher returns would tend to increase the r − g gap, although historically much of the changes in r − g have come about from movements in the return on wealth. During peacetime r has been quite stable, and so has been the r − g gap.
Since the 1970s, the stable and high levels of the rate of return on wealth have coincided with high and rising wealth-to-income ratios (see Piketty and Zucman 2014 and Online Appendix Figure A.7). Together, these facts have meant that the capital share of GDP has increased across advanced economies (Karabarbounis and Neiman 2014). A large part of these high returns, and of the increase in wealth ratios, can be attributed to high capital gains on risky assets, namely housing and equity. Rognlie (2015) argues that house prices have played an important role in the evolution of wealth-to-income ratios in the United States. Kuvshinov and Zimmermann (2018) show that most of the recent increase in the value of listed firms in our cross-country sample is accounted for by higher equity valuations.
These high capital gains in recent decades have allowed the stock of measured wealth to rise without running into diminishing returns. Understanding the drivers behind these long-run trends in returns and valuations seems key to disentangling the underlying causes behind the recent upsurge in wealth, inequality, and the capital share of income.
VIII. Conclusion
In this article we provide an investigation of the long history of advanced economy asset returns for all the major categories of the investable wealth portfolio. Our work brings new stylized facts to light and rigorously documents many broad patterns that have stimulated so much research in core areas of economics and finance over the past two centuries.
The returns to risky assets and risk premiums have been high and stable over the past 150 years. Substantial diversification opportunities exist between risky asset classes, and across countries. Arguably the most surprising result of our study is that long-run returns on housing and equity look remarkably similar. Yet while returns are comparable, residential real estate is less volatile on a national level, opening up new and interesting risk premium puzzles.
Our research speaks directly to the relationship between r, the rate of return on wealth, and g, the growth rate of the economy, that figures prominently in the current debate on inequality. One robust finding in this paper is that r ≫ g: globally, and across most countries, the weighted rate of return on capital was twice as high as the growth rate in the past 150 years.
These and other discoveries can provide a rich agenda for future research. Many issues remain to be studied, among them determining the particular fundamentals that drive the returns on each of the asset classes in typical economies. For now, we hope our introduction of this new compilation of asset return data can provide the evidentiary basis for new lines of exploration in years to come.
En este artículo presentamos una investigación de la larga historia de los rendimientos de los activos de las economías avanzadas para todas las categorías principales de la cartera de riqueza invertible. Nuestro trabajo saca a la luz nuevos hechos estilizados y documenta con rigor muchas pautas generales que han estimulado tantas investigaciones en ámbitos fundamentales de la economía y las finanzas durante los dos últimos siglos.
Los rendimientos de los activos de riesgo y las primas de riesgo han sido elevados y estables en los últimos 150 años. Existen importantes oportunidades de diversificación entre clases de activos de riesgo y entre países. Podría decirse que el resultado más sorprendente de nuestro estudio es que los rendimientos a largo plazo de la vivienda y la renta variable son notablemente similares. Sin embargo, aunque los rendimientos son comparables, el sector inmobiliario residencial es menos volátil a escala nacional, lo que abre nuevos e interesantes rompecabezas sobre las primas de riesgo.
Nuestra investigación aborda directamente la relación entre r, la tasa de rendimiento de la riqueza, y g, la tasa de crecimiento de la economía, que ocupa un lugar destacado en el debate actual sobre la desigualdad. Una conclusión sólida de este trabajo es que r ≫ g: a escala mundial, y en la mayoría de los países, la tasa ponderada de rendimiento del capital fue dos veces superior a la tasa de crecimiento en los últimos 150 años.
Estos y otros descubrimientos pueden proporcionar una rica agenda para futuras investigaciones. Quedan muchas cuestiones por estudiar, entre ellas la determinación de los fundamentos particulares que impulsan los rendimientos de cada una de las clases de activos en las economías típicas. Por el momento, esperamos que la presentación de esta nueva recopilación de datos sobre la rentabilidad de los activos sirva de base probatoria para nuevas líneas de exploración en los próximos años.
An Online Appendix for this article can be found at The Quarterly Journal of Economics online. Data and code replicating tables and figures in this article can be found in Jordà et al. (2019), in the Harvard Dataverse, doi:10.7910/DVN/GGDQGJ.
This article is published and distributed under the terms of the Oxford University Press, Standard Journals Publication Model (https://academic.oup.com/journals/pages/open_access/funder_policies/chorus/standard_publication_model)
No hay comentarios:
Publicar un comentario