The Rate of Return on Real Estate: Long-Run Micro-Level Evidence

Abstract

Real estate—housing in particular—is a less profitable investment in the long run than previously thought. We hand-collect property-level financial data for the institutional real estate portfolios of four large Oxbridge colleges over the period 1901–1983. Gross income yields initially fluctuate around 5%, but then trend downward (upward) for agricultural and residential (commercial) real estate. Long-term real income growth rates are close to zero for all property types. Our findings imply annualized real total returns, net of costs, ranging from approximately 2.3% for residential to 4.5% for agricultural real estate.

Real estate appears to have delivered attractive investment returns over the past few decades (e.g., Favilukis, Ludvigson, and Van Nieuwerburgh 2017; Ghent, Torous, and Valkanov 2019; Giglio et al. 2021).¹ Yet, we possess only a limited understanding of its longer-term track record, especially compared to our knowledge of historical bond and equity returns (e.g., Jorion and Goetzmann 1999; Dimson, Marsh, and Staunton 2002). Recent research by Jordà et al. (2019) suggests that residential real estate has been a stellar investment in a wide range of advanced economies since the late nineteenth century. The authors estimate a geometric (arithmetic) average real net return to housing of 6.6% (7.1%) per year, similar to the performance of equities, with high returns throughout most of the twentieth century.

However, the performance of residential real estate is an inadequate proxy for the performance of direct real estate as a whole. Most obviously, other property types, such as agricultural and commercial real estate, feature prominently in the portfolios of institutional as opposed to individual investors, both today and in the past. Furthermore, estimating the historical performance of direct real estate investments from existing data is fraught with difficulty given the limitations of the available data and the methodological challenges, especially before the late twentieth century. First, available price data generally do not allow for an adequate adjustment for variation in property characteristics. Increases in average prices may thus overstate the capital gains realized by investors if more recently traded properties are of higher average quality. Second, information on the cashflows associated with historical real estate investments is difficult to obtain. When income data exist, they tend to capture contractual instead of realized income, and are not drawn from the same set of properties for which transaction prices are observed. Third, data on the property-level costs associated with real estate ownership are typically not available.

In this paper, we overcome these measurement problems, which are presented more formally in Section 1, by exploiting a unique empirical setting in which we observe transaction prices, rental income, and costs for the same sample of individual properties. Specifically, we construct a data set of the property holdings and transactions starting from the early twentieth century for King’s College and Trinity College in Cambridge and Christ Church and New College in Oxford. These four prominent colleges were among the wealthiest and largest property owners in each university around 1900—and some of them are still economically important real estate investors today (The Guardian 2018). Historically, these colleges have invested in agricultural and commercial real estate as well as housing. Their property portfolios have been professionally managed in pursuit of long-term returns, and were free of major legal or other investment constraints over our sample period.

Our data cover the period 1901–1983 and are hand-collected from the archives of the four Oxbridge colleges. Access to archival records after 1983 is restricted due to the ongoing commercial sensitivity of their data. We source data on acquisitions and disposals from transaction ledgers and on rental income and costs from rent books. We match transactions to the corresponding income records based on property address, tenant name, and other identifying characteristics. In the case of the two Cambridge colleges, the archival sources enable us to assemble the full history of rental income across all property holdings for the entire sample period. For the two Oxford colleges, we collect income data on transacted properties for periods following purchases and preceding sales where possible. Our final data set contains more than 50,000 income observations at the property-year level, and we observe a purchase or sale for nearly 1,800 property-year combinations. At the start of the twentieth century, both the King’s and Trinity portfolios were heavily concentrated in agricultural real estate, which generated more than three-quarters of total gross income. Over the course of the sample period, we document a shift away from agricultural in favor of commercial real estate.

We start the empirical analysis of our novel data set by constructing quality-adjusted rental income indexes over the 1901–1983 period. Our indexes capture the growth in realized income of the property holdings of institutional investors, rather than the growth in average or aggregate contractual rental income in the economy. This distinction is important because marketwide income may increase as new, higher-quality properties are added to the existing stock, while the income for any previously constructed property may not evolve in the same way. We document that real income growth exhibits substantial cyclicality, mirroring inflation and deflation patterns in the U.K. economy. Annualized real growth rates are close to zero for all property types: |$+$|0.3% for agricultural, |$-$|0.3% for commercial, and |$-$|1.0% for residential real estate. These results imply limited capital gains, unless income yields were substantially compressed.

We estimate gross yields associated with matched real estate transactions by dividing the annual gross income generated by a property after a purchase or before a sale by its transaction price. Pooling over all observations, we find a mean gross yield of slightly less than 5%, with more than nine out of ten yields remaining below 8%. Our estimates suggest that the average yields of residential, commercial, and agricultural properties are similar until WWII, fluctuating around 5%. Thereafter they diverge, with yields for agricultural and residential real estate declining, and yields for commercial real estate increasing. Over the final two decades of our sample period, the average yields for agricultural and residential properties are approximately 3%, while the mean yield for commercial real estate exceeds 6%.

Some properties may generate significantly less income than expected, and when transacting some may sell for considerably less (or more) than anticipated. We find substantial cross-sectional variation in yields across individual properties, with an average difference of about two percentage points between the first and third quartiles of the income yield distribution in any given period for each property type. This result highlights the importance of asset-level risks in real estate investment.

Next, we analyze the holding costs associated with real estate investments. Commercial real estate exhibits the lowest average annual expense-to-income ratio (19.4%), especially in the final decades of our study period. Agricultural (28.7%) and residential real estate (32.0%) have higher cost ratios. At the level of individual properties, we document that the impact of costs increases the volatility in property-level net income streams relative to that of gross income streams. Ignoring holding costs can thus lead to a substantial underestimation of the riskiness of real estate investments. In our sample, for a given property-year, the probability of a drop in net income of 10% or more exceeds one in five.

In the case of housing, our estimates of average income growth and net income yields—even when focusing on the same geography and time period—are substantially below those reported in Jordà et al. (2019). The differences are most striking for the decades following WWII. Over the years 1941–1960, we find an annualized real income growth rate of -0.7%, compared to +3.7% in Jordà et al. (2019). Post-1945, 84% of our individual yield observations for residential properties lie below their aggregated average yield series. To generate estimates of annual total returns, including capital gains, we assume that the U.K. price series used in Jordà et al. (2019) is perfectly correlated with true price levels, while allowing for the possibility of a constant (upward or downward) bias. Taking our income index as given, we then identify the average yield series (and the associated price index) that best fits our observed individual yields. Using this approach, we find a geometric average real net total return of 2.3% for U.K. residential real estate over the 1901–1983 period, compared to the estimate of Jordà et al. (2019) of 4.7%. The difference stems not just from lower geometric average income yields (3.0% vs. 4.0%), but also from lower annualized capital gain estimates (-0.7% vs. +0.7%).

Prior estimates of historical income growth rates and returns for agricultural and commercial properties are difficult to find. However, for agricultural real estate, our estimates of income changes and price trends are consistent with the previous research by Lloyd (1992) and Jadevicius et al. (2018). One possible explanation for the greater consistency with earlier studies is that temporal variation in average quality is less of a concern for farmland than for residential real estate. Our results imply an annualized net total return for agricultural real estate of 4.5% over our sample period, boosted by the slightly positive real income growth mentioned before.

For commercial real estate, we are not aware of any income or price index that covers even most of our time period. Nonetheless, our estimates of net yields around 5% for the end of our sample period are broadly consistent with prior findings. By contrast, prior to the 1970s, a large majority of our yield observations fall below the selective (initial contractual) yields reported in the historical study of Scott (1996). While we cannot construct an annual time series of returns for commercial real estate, the combination of negative income growth and substantial income yield expansion implies negative price changes over our time period—mirroring the earlier findings of Wheaton, Baranski, and Templeton (2009) for Manhattan—and thus an annualized total return below the average income yield. Our results therefore suggest a total rate of return for commercial property between those of residential and agricultural real estate.

In sum, our results indicate that direct real estate may be a poorer long-term investment than is suggested by the existing academic literature on housing, either studies covering only more recent time periods or those covering a longer history but combining income and price data from different sources and unable to adjust adequately for quality improvements.

1. Measuring Returns to Real Estate Investments

1.1 Return definitions and decomposition

To understand the nature of direct property investments, we present a decomposition of total returns. We begin by defining the total return to holding property |$i$| between time |$t-1$| and |$t$|⁠, net of the costs associated with property ownership, as:

where |$P_{i,t}$| denotes the market value of property |$i$| at time |$t$|⁠. While |$P_{i,t}$| is not continuously observable, it can be proxied by the transaction price |$P^{*}_{i,t}$| if a transaction takes place at time |$t$|⁠. |$Y_{i,t}$| is gross rental income and |$c_{i,t}$| is the cost-to-income ratio for property |$i$| at time |$t$|⁠, respectively.

We can decompose |$r_{i,t}$| from Equation (1) into its constituent elements, namely, net income yield and capital gain, as follows:

where |$k_{i,t}$| and |$y_{i,t}$| are the capital gain and the gross rental yield of property |$i$| in year |$t$|⁠, respectively. Equation (2) can be expressed in nominal or in real terms; accounting for inflation affects the computation of the capital gain between |$t-1$| and |$t$| (i.e., |$k_{i,t}$|⁠), but not the measurement of the income yield at time |$t$| (i.e., |$y_{i,t}$|⁠).

The total return to the overall market or a specific property type |$\eta$| (i.e., agricultural, commercial, or residential), |$r_{\eta,t}$|⁠, then can be defined as the aggregated total return over all assets |$i=1,2,...,N$| that are part of |$\eta$| at time |$t-1$|⁠. We then have that

Because of infrequent trading and temporal variation in average property quality, capital gains |$k_{\eta,t}$| may be difficult to estimate without bias from transaction prices (cf. infra). Therefore, it is useful to rewrite Equation (2) as follows, using |$P_{i,t}=\frac{Y_{i,t}}{y_{i,t}}$|⁠:

where |$g_{i,t}$| denotes the property-level income growth rate between |$t-1$| and |$t$|⁠. Assuming that the yields computed from transaction data, namely, |${y^{*}_{i,t}}\equiv \frac{Y_{i,t}}{P^*_{i,t}}$|⁠, are representative for the asset class in any period, we can compute a property type’s net total return from the observed aggregate income growth rates, changes in gross income yields, and aggregate cost ratios as follows:

Equations (4) and (5) highlight the importance of income growth in long-term real estate returns. Absent (future) income growth, price increases will imply higher capital gains today, but lower income yields going forward. Absent relative price (i.e., yield) changes, capital gains will equal real income growth rates. Given that yields cannot increase or decrease ad infinitum, long-run average income growth rates must necessarily be an important determinant of long-run average capital gains.

1.2 Measurement problems in prior literature

The existing literature on the long-term financial characteristics of real estate investments mainly focuses on housing as an asset class. In addition, much of the existing work estimates a time series of average capital gains |$k_{\eta,t}$|⁠, while time series of income yields |$y_{\eta,t}$| or total returns |$r_{\eta,t}$| are estimated less frequently. Guided by the decomposition of the net total returns to individual property investments shown in the previous subsection, we can summarize the measurement problems faced by prior studies estimating the long-run performance of real estate as follows.

1.2.1 Capital gains

Many studies document the historical evolution of aggregate house price indexes, based on average transaction values |$\overline{P^{*}_{t}}$| observed in different housing markets. Important contributions include Eichholtz (1997) for Amsterdam, Shiller (2000) for 20 cities in the United States, and Knoll, Schularick, and Steger (2017) for 14 different countries. However, a time series of changes in average observed transaction prices |$\frac{\overline{P^{*}_{t}}}{\overline{P^{*}_{t-1}}}$| misestimates the average capital gains realized by investors if it does not adequately control for changes in the quality composition of the traded real estate stock. In particular, if new properties are of higher average quality than existing ones, then average transaction prices will increase at a rate that exceeds the capital gains of existing investors. This is a well-known problem encountered in empirical studies of housing price trends. For example, the U.K. house price index of Knoll, Schularick, and Steger (2017) “does not control for quality changes prior to 1969” (appendix p. 114); more generally, the authors acknowledge that “accurate measurement of quality-adjustments remains a challenge” (p. 342). Next, investors will only realize capital gains in line with a quality-adjusted price index if they maintain their property; however, maintenance expenditures are not always taken into account in the literature. Finally, there exists a “superstar city bias” in that many historical studies—even of “national” housing prices—focus on capitals and other large cities (Dimson, Marsh, and Staunton 2018), which are known to have had a higher-than-average rate of price appreciation historically (Gyourko, Mayer, and Sinai 2013).

1.2.2 Income yields

Modern financial institutions invest in real estate expressly in pursuit of high rental yields (Hudson-Wilson et al. 2005). However, existing empirical studies on the investment performance of real estate largely ignore the rental yield component of total returns. Rental yields are absent from prior research in part because data on the cashflows from real estate investments are difficult to obtain. Unlike transaction prices, rental income observations are not systematically or centrally recorded. Furthermore, existing data tend to capture contractual rental income, which can be significantly higher than realized income due to rent arrears and temporary voids, and thus lead to overstated rental yields. The ratio of contractual rents to market rents also may be highest at the time of establishing the contract, which is typically when it is observed.

Where income data are available for property samples, transaction prices are not typically observed for the same properties. As a result, researchers often combine income and price data from different sources to estimate yields and total returns. For example, Nicholas and Scherbina (2013) compute total returns to Manhattan real estate investments between 1920 and 1939 by adding survey data on contractual rents from 54 income-producing local properties to their price index estimated over a different property sample. Brounen et al. (2014) combine the long-run Amsterdam price index from Eichholtz (1997) with the rent index from a different set of properties constructed by Eichholtz, Straetmans, and Theebe (2012) to estimate total returns. Most recently, Jordà et al. (2019) provide total return estimates on housing for 14 countries by combining price indexes from Knoll, Schularick, and Steger (2017) and rental income indexes estimated from aggregate national statistics in Knoll (2017). However, all those methods introduce a measurement error in the resultant time series of income yield estimates if the underlying price and rent observations are obtained from properties with different (quality) characteristics (Eichholtz, Korevaar, and Lindenthal 2018). Moreover, using historical income and price change estimates to extrapolate backward from a contemporary rent-to-price ratio over long periods of time, as in Brounen et al. (2014) and Jordà et al. (2019), runs the risk of compounding measurement errors over time (Dimson, Marsh, and Staunton 2018).

Finally, the data underlying existing long-run studies typically do not account for the actual asset-level costs of property ownership and management which have to be borne by the investor and which can materially affect the net income generated from direct real estate investment.

1.2.3 Income growth

An accurate measurement of historical rental income growth rates can inform the analysis of long-term real estate returns. As mentioned before in the context of yield estimation, some prior studies construct market-level housing rent indexes. However, as with price indexes, controlling for changes in the quality mix of properties over time remains a challenge. The issue may be particularly relevant when indexes are based on aggregate national statistics, such as when Knoll (2017) relies on the rent component of cost-of-living and consumer price indexes. In a recent contribution, Eichholtz, Korevaar, and Lindenthal (2018) explicitly tackle the issue of quality adjustments using new historical data and conclude that “most of the increase in housing expenditure that did occur is attributable to increasing housing quality rather than rising rent.”²

1.2.4 Idiosyncratic risks

Most research focuses on aggregate capital gains or total returns, mainly in the housing market. However, the return on an individual property |$r_{i,t}$| may be substantially above or below the aggregate return for its property type |$r_{\eta,t}$| for several reasons. First, “transaction-specific risk” is nonnegligible: if property |$i$| transacts in year |$t$|⁠, this may happen at a particularly low or high price |$P^{*}_{i,t}$| that is different from the (unobservable) |$P_{i,t}$|—an idea dating back to at least Shiller and Case (1987), and discussed more recently in Lovo and Spaenjers (2018), Giacoletti (2021), and Sagi (2021). Second, for any given property and in any given year, the gross income |$Y_{i,t}$| may be lower than anticipated (i.e., contracted) or costs |$c_{i,t}$| may exceed expectations for that year. This last source of idiosyncratic risk—and in particular the possibility that annual returns may be volatile because of variation in expenses—has not been explored in prior literature. For example, the discussion of idiosyncratic risks as a potential explanation for the “housing risk premium puzzle” in Jordà, Schularick, and Taylor (2019) focuses only on individual house price volatility.

1.2.5 Property types other than housing

Residential real estate accounts for almost three-quarters of the US

230 trillion global market (Tostevin 2017). However, only one third of that share is investable, compared to two-thirds of commercial.³ As a result, an analysis of housing alone ignores the substantial opportunity set of commercial and agricultural property available to institutional investors. It is therefore problematic to draw inferences about the investment performance of real estate from the performance of residential property alone. To date, only a limited number of long-run studies focus on property types other than housing. Examples include research on historical trends of U.K. agricultural income (Lloyd 1992) or prices (Jadevicius et al. 2018), commercial property prices in Manhattan (Wheaton, Baranski, and Templeton 2009), or commercial real estate income yields in the United Kingdom (Scott 1996). These studies provide useful benchmarks for different dimensions of our findings.

1.3 This paper

In our hand-collected data set, described in Section 3, we directly observe income |$Y_{i,t}$|⁠, income growth |$g_{i,t}$|⁠, and costs |$c_{i,t} Y_{i,t}$| across individual properties of different types |$\eta$|⁠. We can also compute the gross income yield |$y^{*}_{i,t}$| if property |$i$| transacts in year |$t$|⁠. These data enable us to shed more light on historical income growth, yields, and ultimately total returns for different property types. In addition, we are able to document cross-sectional variation at the property level in yields and in net income growth rates.

2. Oxbridge Colleges as Real Estate Investors

Our data set draws on the real estate investments held by some of the most prominent Oxford and Cambridge colleges. U.K. institutional investors have a long record of investing in real estate, and the oldest Oxbridge college endowments have held property for at least five centuries. For example, King Henry VIII founded Trinity College, Cambridge, and Christ Church, Oxford, in 1546 and conferred on both colleges a diversified agricultural real estate portfolio. At the start of the twentieth century, their portfolios consisted almost exclusively of real estate (Chambers, Dimson, and Foo 2015a). This stands in contrast to the largest U.S. university endowments, which allocated only 10% of their assets to direct real estate in the early twentieth century (Goetzmann, Griswold, and Yung-Fang 2010); U.S. institutional investors in general did not begin to make a substantial allocation to real estate until the 1970s and 1980s (Eagle 2013). Notwithstanding diversification into stocks and bonds in the twentieth century, the oldest and wealthiest of the Oxbridge colleges still allocate over 40% of their endowment to real estate today (Cambridge Associates 2018).

The starting point of our empirical approach was to identify substantial investors in U.K. real estate in 1900 with a diversified portfolio of properties and, crucially, with comprehensive and accessible archival records of their rents and transactions. Our resultant sample comprises a small set of investors. Yet, there are reasons to believe that the colleges we study are sufficiently representative of the population of wealthy institutional investors they are drawn from.

First, the biggest Oxbridge colleges were able to devote substantial resources to the professional management of their property portfolios—and their endowments more generally—over our sample period. This is particularly apparent in the cases of Trinity and King’s College in Cambridge (Nicholas 1960; Neild 2008; Chambers, Dimson, and Foo 2015a). The senior bursars (or treasurers) of the colleges—equivalent to the roles of chief financial officer and chief investment officer combined into one—were typically academics without much experience of finance and investment. (John Maynard Keynes, the bursar at King’s from 1921 to 1946, was very much the exception being an economist and accomplished investor (Chambers, Dimson, and Foo 2015b).) Nonetheless, they brought an open-minded intellect to managing the endowment and the role was full-time. In addition, the bursars received considerable support from two sources. First, advice and expertise came from alumni with relevant finance and real estate experience who sat on the estates committees (responsible for property management) and the finance committees (responsible for overall financial and investment policy). For example, Anthony de Rothschild, head of the family merchant bank in London, sat on the Trinity finance committee from 1930 onward (Neild 2008). These committees met regularly throughout the academic year and frequently produced memos on investment strategy as well as reviews of investment performance. Second, the bursars were professionally advised by specialist estate agents, such as Bidwells, Savills, and Smith Woolley. (In some cases, colleges also appointed a dedicated estates bursar with professional property qualifications and experience.) As to the question of whether academic institutions are better investors than nonacademic institutions, the evidence is mixed. Some studies have argued that the wealthiest U.S. university endowments are better investors (Lerner, Schoar, and Wang 2008; Chambers, Dimson, and Kaffe 2020), while others claim the opposite (Ennis 2021).

Second, these colleges faced no major constraints on their ability to manage their investment properties—the subject of our study—as distinct from their operating properties. As such, they were no different to any other institutional investor with an allocation to direct property. More specifically, from the late 1850s onward, colleges were allowed to sell their originally endowed real estate and reinvest in new properties, and were able also to charge market rents (Dunbabin 1975; Neild 2008).

Third, while the college portfolios had been historically heavily reliant on agricultural property as a result of their legacy endowments, the regional distribution of their portfolios in 1901 does not suggest a significant geographical bias in the quality of farmland as measured by crop yields (Collins 2000). Moreover, the adverse impact of a severe agricultural depression in the latter half of the nineteenth century had made clear the need to manage their portfolio more actively. By the early part of our sample period, therefore, the colleges were already investing to improve the quality of tenants and raise rents, and had begun disposing of farms with relatively poor rental growth prospects. They were also now diversifying into commercial property for the first time. More generally, the endowments were clearly being managed in the pursuit of long-term investment returns.

3. Income and Transaction Data

3.1 Data collection

We study King’s College (founded in 1441) and Trinity College (founded in 1546), Cambridge, and Christ Church (also founded in 1546) and New College (founded in 1379), Oxford. These colleges were among the oldest and wealthiest Oxbridge colleges at the beginning of the twentieth century (Dunbabin 1975).

Our data collection starts in 1901 and continues until 1983 (except for New College, where we only have data until 1955). Access to archival records after 1983 is restricted by the colleges, due to the ongoing commercial sensitivity of their data, given the long-term relationships they have with some of their tenants. Nonetheless, our sample period is sufficiently long to enable us to draw meaningful comparisons with earlier research.

Our focus is on the investment properties held in college endowments. We ignore operational properties outside the endowments, which are typically not for sale and are not let at market rents. Similarly, we do not consider furnished lettings and lodging houses in Cambridge and Oxford, as they were used to house faculty or students. We also exclude nonstandard rent types, typically associated with very small sums, such as rentcharges and wayleaves.

We compile an unbalanced panel data set of property-year observations on rental income received, costs incurred, and transaction values realized, alongside a number of property and transaction characteristics, as follows.

For King’s College, Cambridge, we collect annual property-level realized rental income and costs incurred over the years 1901–1983 from the annual volumes of the so-called “Mundum Books.” Transactions of properties over the same period are found in the “Ledger Books.” We record transaction type (purchase or sale), transaction amount, and year. These records further allow us to identify partial transactions (part of a property) and portfolio sales (more than one property), as well as instances where the use of the property has changed—often agricultural land sold with approval for residential development. Where available, we collect information on location, size, and other characteristics of the property. We manually match income and cost records with transaction records based on the common property name.⁴

Figure 1 presents an example of a matched transaction for King’s College. Panels A and B show the income and cost record for a farm property called “Middle Cliston” located in the village of Sampford Courtenay in the county of Devon, in the southwest of England. In the financial year 1926, the property generated a rental income of £152 in two semiannual payments, while the costs (including for “ironmongery” and a “carpenter”) amounted to slightly more than £19 in total. Panel C shows the part of the document recording the transaction for the same property, which was sold in 1927 for £2,552.

For Trinity College, Cambridge, we collect annual realized property-level rental income and costs incurred from the “Rent Books” over the period 1901–1983. (A missing volume in the archives means we lack the records for some residential properties for the period 1966–1971. We impute income and costs for properties with records in 1965 and 1972 through geometric interpolation.) Unlike at King’s, annual income and cost data are reported by tenant name, instead of property name. When several tenants occupy a property, we aggregate income across all tenants in the property in a given year. We collect data on transactions over the same period from the summaries in the annual college accounts and the “Sealed Books.” Transactions are reported by property name. Matching tenant-level income data with property-level transaction data is challenging, particularly in the earlier part of our sample period, as there is no common identifier between the property on which the college receives rent from a tenant and the property transacted. In such cases, we match the Trinity income and transaction records by identifying unique property characteristics, such as exact address and property type information.

In the case of Christ Church and New College in Oxford, there are some gaps in the archival records, particularly with respect to rents, because of volumes being destroyed or lost. For Christ Church (New College), we collect data on transactions over the period 1901–1983 (1901–1955).⁵ We then collect income and cost data for the year of the transaction and for the 2 years before (after) a sale (purchase). For New College, we obtain information on contractual income where records for realized income are missing (less than 20% of all observations for the college). Data on transactions and annual property-level income are obtained from the sources outlined in Internet Appendix A. Properties are identified by name in all sources, allowing us to match income and cost as well as transaction records. Property and transaction characteristics are collected from all sources as available.

3.2 Data sets and descriptive statistics

The data collection described above produces two data sets. The first is a database of 52,761 annual property-level observations on income, cost and property characteristics, mainly based on records from King’s and Trinity College.⁶ The unit of observation is a unique property-year combination, i.e., property |$i$| in year |$t$|⁠. The data set covers 3,046 different properties.

The second is a database of 1,769 matches between our property-level income data and transaction records, which is used in our estimation and analysis of income yields. This total does not include transactions of parts of properties; however, we keep track of them when estimating income indexes so that we can exclude income changes over years in which such partial transactions occurred. Our matched database includes income data as well as corresponding property and transaction characteristics. Here, the unit of observation is a property-transaction combination, that is, the purchase or sale of property |$i$| in transaction |$x$|⁠. (Every transaction is associated with a certain year |$t$|⁠.) Since some transactions involved more than one property identifiable in the rental income records, our final matching of property records corresponds to 1,359 distinct transactions.

Table 1 summarizes for the income database the number of property-year income observations, and the number of distinct properties that these observations relate to, by decade (panel A), college and property type (panel B), and region (panel C). Our data are spread quite evenly over time, although we see a decline in the number of properties in later decades. Based on the qualitative descriptions of the dominant use provided in the archival records, we classify properties into one of four main types. “Agricultural” refers to land used for farming and represents approximately 35% of the sample. “Commercial” refers to any property let to a retail, office, industrial, or other commercial business and represents 17% of our income observations. “Residential” refers to any residential property or building and contributes about 39% of the data. “Other”—by far the smallest category—refers to schools, government buildings, gardens, etc. The Oxford colleges have lower numbers of observations due to the restrictions we face on data availability and sampling.

Panel C of Table 1 shows that the Oxbridge colleges were diversified geographically, although the largest portfolio shares were located in the more prosperous East and South of England (including London). Figure 2 shows a more granular breakdown of the portfolio holdings of King’s and Trinity College, Cambridge by aggregate income in the counties of England, Scotland, and Wales, at the beginning and end of our sample period—in 1901 (panel A) and in 1983 (panel B), respectively.

Table 2 reports on the composition of our sample of properties where we are able to match rental income to a transaction. Although matched property sales outnumber purchases by a ratio of about 2:1, we will see in Table 3 that the average real purchase price is about three times the average sale price. Panels A–C summarize the distribution of matched property-transaction records by purchase and sale over time, by college, and by property type. The final column in each panel shows the number of distinct transactions—some of which involved more than one property identified in the rental income records—corresponding to each row. We observe the highest trading volume in agricultural real estate, particularly in terms of sales. In panel D, we show the breakdown of matched observations between transactions of a single property and those of a portfolio of multiple properties.

Table 3 provides summary statistics on nominal and real income and transaction prices over time. We use inflation data from Dimson, Marsh, and Staunton (2020) to convert nominal values into real (beginning of 2020) terms. In panel A, we report for each decade the number of nonmissing (gross) income observations, together with means and medians. In real terms, mean income decreased over the initial years of the twentieth century, but increased substantially in later decades. However, median real income only bottomed out in the 1960s. The substantial increases in mean and median income in the later part of our time frame can largely be explained by the increasing relative importance of commercial real estate in the college portfolios (cf. infra). In panel B, we show for each decade the mean and median purchase and sale price levels. In this panel, the unit of observation is a transaction, so that each portfolio transaction is only counted once. Mean purchase prices are substantially higher than sale prices in all decades after the first one. Although mean prices are higher at the end of our sample period than at the beginning, there exists no monotonic trend over time.

3.3 The evolution of real estate portfolios

We can analyze how the nature of institutional real estate investment has changed over time by charting the evolution of the King’s and Trinity property portfolios, for which we have a near-complete history of holdings and associated rental income. Figure 3 graphs the evolution of the two portfolios. Panels A and B show the time series of the number of income-generating properties by property type for both colleges. Panels C and D display the evolution of real (inflated to year 2020 British pounds) income along the same dimensions. Taken together, the two Cambridge colleges always had several hundred properties in their portfolio. By the end of our sample period, their aggregated income amounted to about ten million pounds in year 2020 terms.

The portfolios of these institutions have always encompassed all three property types. At the start of the twentieth century, both colleges were heavily concentrated in agricultural real estate, which generated more than three-quarters of total gross income. This allocation reflects the nature of their original endowments received centuries earlier and the low turnover during the intervening period. Yet, from the little that is known of other institutional property portfolios at the very start of the twentieth century, a heavy portfolio allocation toward agricultural real estate was not unusual (Dunbabin 1975). Subsequently, over our sample period there was a significant shift in the portfolio holdings of the two colleges—in terms of both the number of properties and total rental income—away from agricultural and toward commercial real estate. In later decades, commercial real estate becomes the most important property type for both institutions. Its growing relevance is most striking when studying the evolution of income rather than the number of properties, as average income is higher for a commercial property than for a residential dwelling. Our findings underscore the need to include property types other than housing in order to accurately estimate the long-run performance of the direct real estate asset class as viewed by institutional investors.

4. Estimates of Income Growth and Income Yields

4.1 Gross income growth over time

To estimate quality-controlled (gross) income indexes over our sample period, we proceed as follows. In every year |$t$| and for each property type, we consider all Cambridge college properties that are present in our income database both in year |$t-1$| and in year |$t$|⁠. To avoid mismeasurement of income trends due to properties that have just entered or are about to exit the portfolio, we exclude from the analysis any properties for which we observe a (nonpartial or partial) transaction between |$t-2$| and |$t+1$|⁠, and properties that first appear in the data set in |$t-2$| or |$t-1$| or last appear in the data set in |$t$| or |$t+1$|⁠. Next, we compute for every year |$t$| the percentage change in aggregate income. Finally, we chain-link the estimated time-series of income growth rates. Our method thus explicitly controls for changes in the quality composition of the portfolio of properties over time, in a way similar to repeat-sales regressions in the estimation of capital gains on properties. Figure 4 presents the resultant deflated indexes for each of the three main property types alongside annual inflation rates.

The figure shows substantial time-series and cross-sectional variation in real income growth. Real rental income decreased dramatically for all property types between 1913 and 1920 due to the high inflation during the war and its immediate aftermath; it then rebounded in the early 1920s as a result of the subsequent price deflation induced by a policy of returning the pound sterling to the gold standard. Income for all property types decreased in the second half of the 1930s, as inflation started to rise again. Between 1940 and 1960, there is very little real income growth, partially reflecting the introduction of rent controls (Knoll 2017). After 1960, agricultural and commercial income indexes increase substantially, while the residential income index trends downward.

Table 4 presents annualized real income growth rates over our complete sample period 1901–1983, and for four different subperiods: 1901–1920, 1921–1940, 1941–1960, and 1961–1983. Considering the full sample period, we estimate that real income growth is close to zero for all property types. The annualized real growth rates based on the indexes shown in Figure 4 are +0.3% for agricultural, -0.3% for commercial, and -1.0% for residential real estate. The latter result implies that the colleges could only have achieved capital gains in excess of inflation on their commercial and residential real estate investments in a situation where gross income yields declined over time.

4.2 Gross income yields

Based on our matched rental income and transaction data, we estimate gross income yields. We only consider transaction values exceeding £100 in year 1901 terms (approximately £6,600 at the start of 2020) and ignore partial transactions. Additionally, we exclude cases in which the income only constitutes ground rents (i.e., rents paid under the terms of a lease by the owner of a building to the owner of the land), transactions associated with a change of use (e.g., agricultural land being sold for development), and properties classified under the “other” category. (In Internet Appendix B, we plot the yields for all these dropped transactions, which are relatively low on average.) For a property bought (sold) in year |$t$|⁠, we use the maximum real income generated by the property over the calendar years |$t$| until |$t+2$| (⁠|$t-2$| until |$t$|⁠). We exclude cases in which this maximum equals zero. We consider these 2-year windows before (after) a disposition (acquisition) to minimize the effect of transaction-related temporary voids on income and thus estimate the income generating capacity of a given property under normal conditions. Next, we divide real income by the real transaction price.⁷ In the case of portfolio transactions, we only observe a single transaction price for the entire portfolio. To compute yields on portfolio transactions, we aggregate income over all properties reported as being bought or sold in the same transaction. We classify the transaction under a single property type (agricultural, commercial, or residential) according to which property type generates the largest share of total income. We truncate the lowest and highest 2.5% of yield estimates.

Figure 5 shows a scatterplot of all resultant individual yield estimates for agricultural, commercial, and residential real estate. Table 5 presents summary statistics by property type for gross income yields, estimated over the same subperiods shown in Table 4 and over the full 1901–1983 period.

Pooling all observations in Figure 5, we compute a mean gross yield of 4.7%. More than nine out of ten yields are below 8%. However, there is some variation in mean yields across property types and over time. In Table 5, agricultural and residential yield estimates hover around 5% before WWII and decline thereafter. Commercial yields trend upward in each successive subperiod. Toward the end of our sample period, the contrast between the low observed yields for agricultural (mean of 2.8% over 1961–1983) and for residential real estate (mean of 3.2% over 1961–1983) and the high yields for commercial real estate (mean of 6.5% over 1961–1983) is particularly striking.

At the same time, Figure 5 and Table 5 highlight the cross-sectional variation in realized yields across individual real estate assets, even within a certain time period and property type. Table 5 exhibits an average difference of about 2 percentage points between the first and the third quartiles of income yields for any 20-year period and property type. This finding underlines the importance of asset-level risks in real estate investments. Transaction prices can be lower or higher than anticipated, as studied in more depth by Giacoletti (2021) and Sagi (2021). Moreover, gross yields can be (temporarily) depressed by rental voids or low-quality tenants.

4.3 The impact of costs

We now analyze the extent to which costs—mainly repairs and improvements, but also property taxes and rates, insurance, agency costs, etc.—reduce realized income yields.⁸ To assess the importance of costs, we compare them to realized income (in the following year) aggregated across the two Cambridge colleges. Similar to our income index estimation, we exclude from the analysis any properties for which we observe a (nonpartial or partial) transaction between |$t-1$| and |$t+1$|⁠, and properties that first appear in the data set in |$t-1$| or |$t$| or last appear in the data set in |$t$| or |$t+1$|⁠.

Figure 6 displays the resultant time series by property type. The figure shows that cost ratios are typically around 20%–30%. The downward trend in costs for commercial real estate near the end of our time frame could be related to the greater incidence of contract types where the tenant is responsible for maintenance and repairs. The spikes in residential real estate cost-to-income series near the end of our sample period are due to the impact of a small number of substantial renovations.

Table 6 shows the associated time-series averages. Over the whole sample period, commercial real estate is associated with the lowest average cost ratio (19.4%), in particular in the final decades of our study period. Agricultural (28.7%) and especially residential real estate (32.0%) have higher relative costs, driving a larger wedge between gross and net yields for these property types. Compared to modern estimates for the United States, these estimates are arguably on the low side. Eisfeldt and Demers (2018) report an average contribution of expenses to gross yields of 41% for single-family rentals over the period 1986–2014. For twenty-first century commercial real estate, NCREIF data show an average expense-to-income ratio of about 35%.

Figure 6 focuses on aggregate costs. The substantial variation across properties and over time in the impact of costs means that property-level net income streams will be much more volatile than gross income streams. This finding is illustrated in Figure 7, which shows the annual distribution of property-level gross and net income changes over the prior year across five categories: a decrease of 10% or more, a decrease of less than 10%, no change, an increase of less than 10%, or an increase of 10% or more. Panel A shows that sharp decreases or increases in gross income are relatively uncommon. (We find that income drops to zero in about 0.5% of all observations.) Panel B tells a completely different story once costs are taken into account. When pooling data across years, we estimate that the probability of a property-level decrease in net income of 10% or more exceeds 20%. Ignoring costs thus leads to a substantial underestimation of the volatility of real estate income streams.

5. Discussion and Total Return Estimates

In this section, we discuss our findings on historical income growth rates and income yields for residential and nonresidential real estate in the context of the existing evidence. We also provide total return estimates and, where possible, benchmark them against prior research.

5.1 Residential real estate

Most of the existing empirical work on the performance of real estate investments relates to the history of the housing market. We directly compare our index of residential rental income to the rental index in Jordà et al. (2019) based on Knoll (2017) and to the quality-adjusted real rent index for London estimated recently by Eichholtz, Korevaar, and Lindenthal (2018) by plotting the different series in panel A of Figure 8 for all available years starting in 1901. Note that Eichholtz, Korevaar, and Lindenthal (2018) use a methodology similar to ours for most of the overlapping time period, relying on repeated individual rent observations taken from the archives of real estate investors. By contrast, the estimates used in Jordà et al. (2019) are constructed using aggregate data on average rents and on the rent component in consumer price indexes that do not adjust for variation in quality. Importantly, the primary data are particularly thin for the 1939–1954 period; Knoll (2017) even notes that “to the best of [her] knowledge, no data on rents exist between 1946 and 1954” (p. 247).⁹

In Table 4, we report an estimated decline of 1.0% per year in real income for U.K. residential property over the 1901–1983 period. By contrast, the index of Jordà et al. (2019),grows by 0.9% per year on average. Our estimate is closer to that of Eichholtz, Korevaar, and Lindenthal (2018), who compute an annualized decline in real housing rents for London of 0.6% per year over the same period. Figure 8 shows a substantial gap in estimates for the period between the start of WWII and the early 1960s. Panel A of Table 7 shows that, over the years 1941–1960, we find an annualized real income growth rate of -0.7%, compared to +3.7% in Jordà et al. (2019). The difference in income growth estimates between our results and those presented in Jordà et al. (2019) underscores the importance of controlling for variation in housing characteristics when estimating long-run rental growth figures for residential real estate.

The large gap in income growth estimates between our results and Jordà et al. (2019) for the post-WWII years coincides with substantial differences in the income yield estimates.¹⁰ In panel B of Figure 8, we graph our individual yield observations, converted to net yields by assuming the average cost ratio reported in Table 6. We also plot the average yield series of Jordà et al. (2019). We find that 74% of our individual yield observations are below their series; this fraction increases to 84% for the post-1945 period.¹¹ Overall, Jordà et al. (2019) estimate a (geometric and arithmetic) average aggregate net yield of 4.0% over our period, while we reported an arithmetic average gross yield of 4.3% over all individual residential property yield observations in Table 5 and a cost-to-income ratio close to one-third in Table 6, implying much lower net numbers.

Our finding that residential income declined by half in real terms over our sample period, while income yields decreased less, implies negative capital gains—and therefore average total returns lower than average income yields. While the relatively high frequency of our rental data allow a precise estimation of annual income growth, our transaction price data do not permit the direct estimation of an annual capital gains series (i.e., |$k_{\eta,t}$| in Equation (3)) with sufficient precision. However, using Equations (4) and (5), we demonstrated that the time series of income growth and average yield estimates can help us generate an estimate of historical capital gains. Having already reported our income index, the remaining challenge is to extract from our sample of individual yield observations estimates of the annual series of average yields (i.e., |$\overline{y^{*}_{t}}$|⁠), and thereby to construct an annual series of all return components that can be compared to the results of Jordà et al. (2019). Since we only observe a limited number of transaction data points in each time period, we discipline the estimation of the average yield series as follows. We assume that the estimated U.K. housing price index of Jordà et al. (2019), which is based on Knoll, Schularick, and Steger (2017), is perfectly correlated with the true underlying price index over our sample period, but that annual changes in their index may over- or underestimate true capital gains by some constant parameter.¹² Taking our income index as given, we can then find the average yield series that best fits our observed individual yields. We do so through a simple regression model in which we estimate two coefficients: one related to the relative average level of income to prices, and one related to the degree to which the benchmark price index over- or underestimates annual price changes. We provide more details on our procedure in Internet Appendix C.

Panel B of Figure 8 plots the resultant fitted series of average income yields. It tracks that of Jordà et al. (2019) rather closely until the start of WWII, but then diverges substantially. In panel C, we benchmark the estimated capital gains series against each other. The price index values of Knoll, Schularick, and Steger (2017) used in Jordà et al. (2019) imply a positive average real capital gain over the 1901–1983 period. (We geometrically interpolate their missing price index values over the WWII period.) By contrast, we find that real price levels have gone down over our sample period. The same panel also shows our resultant total returns index, and compares it to that of Jordà et al. (2019). In panel B of Table 7, we report the associated mean statistics. In sum, we find lower geometric average capital gains (-0.7% vs. +0.7%) and net income yields (3.0% vs. 4.0%) than Jordà et al. (2019), leading to an annualized net total return estimate of 2.3%, compared to their 4.7%. This substantial difference in the two headline results likely arises from the greater precision of income growth and yield estimates enabled by our granular property-level data set. Also, our empirical strategy avoids the compounding of errors over time that is inherent in the backward extrapolation, starting from contemporary yields, employed by Jordà et al. (2019).

Finally, we note that, in the study of Jordà et al. (2019), the United Kingdom is associated with some of the lowest estimates—out of a sample of 16 countries—of the net returns to housing. For the period 1901–1983, averaging across the country-level geometric mean net return estimates gives a value of 6.3%, which is substantially above the U.K. estimate. Therefore, we believe that the discrepancies that we document are unlikely to be due to U.K.-specific measurement error in Jordà et al. (2019).¹³

5.2 Agricultural real estate

Estimates of historical income growth rates are more difficult to find for agricultural property. Nonetheless, we compare our income index to the agricultural land rent index constructed from survey data in the little-known study of Lloyd (1992). We graph the results in panel A of Figure 9. Overall, the two series appear quite consistent. This finding provides additional comfort regarding the representativeness of our agricultural income data.

In panel B of Figure 9, we plot our individual yield observations, but now converted to net yields by applying the sample-period-mean cost ratio of 28.7% (cf. Table 6). Although we lack a comparable series for our entire sample period, our estimated yields for the early 1980s are broadly consistent with the earliest numbers reported by IPD.

To construct a fitted net yield index following the same approach outlined above, we make use of the farmland price index in Jadevicius et al. (2018), which heavily relies on a farm auction price series constructed at Oxford University. Interestingly, we do not need to (significantly) adjust their annualized capital gain estimate downward to obtain the best fit of our observed yields. Their estimate of 1.1% is identical to our own figure reported in panel B of Table 7. As a result, the two time series of prices in panel C of Figure 9 virtually coincide.

Thus, our estimates seem to agree with the (limited) previous research on agricultural real estate in capturing income and price trends. Most likely, temporal variation in average property quality is less of a concern for agricultural than for residential real estate. More importantly, our research is novel in that we can produce estimates of agricultural yields and thus also total returns. From the fitted index in panel B of Figure 9, we estimate the average income yield on agricultural real estate over our sample period to be 3.4%. Panel C shows the total returns index that results from our analysis. Panel B of Table 7 reports an annualized net total return of 4.5% over our sample period, substantially exceeding that of residential real estate.

5.3 Commercial real estate

In the case of commercial real estate, we are not aware of any income or continuous price index that covers even most of our time period. Figure 10 displays our individual yields, again converted to net numbers in the same way as before, and compares them to the annual average yields reported for shops and offices in Scott (1996). For the period 1971–1983, we also show the yield estimates from IPD.

Our estimated net yields of around 5% at the end of our sample period are broadly consistent both with Scott (1996) and with the earliest-available data on U.K. commercial property provided by IPD. However, prior to 1970, a large majority of our yield observations fall below those reported in Scott (1996). A possible explanation for this difference is that Scott (1996) reports initial contractual yields on property investments, which are arguably an upper bound estimate of actual income yields realized later in the holding period.¹⁴ In addition, some of the data in Scott (1996) come from information collected on investment purchases by financial institutions and from investment property newspaper advertisements, both of which may overstate the yields achieved by the average institutional investor.

While we cannot construct an annual time series of commercial real estate returns, the combination of negative income growth documented previously and substantial income yield expansion implies negative price changes over our time period. This result is consistent with earlier findings of Wheaton, Baranski, and Templeton (2009), who estimated that Manhattan office property values were substantially lower in 1999 than in 1899 (based on a small sample of repeat-sales transactions). We can therefore consider the average net income yield of about 4.5% reported in Table 5 to be an upper bound estimate of total returns. Our results thus suggest a performance in between that of residential and agricultural real estate.¹⁵

6. Conclusion

In this paper, we construct a unique data set from the archival records of a group of important U.K. institutional real estate investors—the endowments of four large Oxbridge Colleges. The data comprise granular information on realized rental income, costs, and transaction prices at the level of individual properties over the period 1901–1983.

The evolution of our sample of direct property portfolios over this time period is indicative of the greater importance of nonresidential compared to residential real estate for U.K. institutional investors. To better understand the past performance of direct real estate as an asset class, we thus also need to examine agricultural and commercial real estate. Our data set allows us to estimate not only realized income growth over time but also income yields, cost-to-income ratios, and net total returns for all three property types.

The main empirical results can be summarized as follows. First, long-term real rental income growth is close to zero for all major property types. Second, average gross yields fluctuate around 5% for most of the study period. Significant yield compression in agricultural and residential real estate occurred during the last decades of our sample period, while commercial yields displayed the opposite trend. Third, operating costs lower gross yields by 20%–30%, and significantly increase the volatility of the net income from real estate. Fourth, our estimates of income and yield dynamics imply very limited (or even negative) long-run capital gains, and annualized real net total returns of between 2.3% (residential) and 4.5% (agricultural).

Overall, our results suggest that the investment performance of direct real estate assets in the long run is less attractive than previously thought.

Authors have furnished an Internet Appendix, which is available on the Oxford University Press Web site next to the link to the final published paper online.

Acknowledgement

We thank Aubrey Adams, Stephen Broadberry, Roland Bull, Judith Curthoys, Elroy Dimson, Rui Esteves, William Goetzmann, Rowena Gray, Andrew Karolyi, Matthijs Korevaar, Roman Kräussl, Dmitry Kuvshinov, James Lawrie, Ryan Lewis, Thies Lindenthal, Robert Margo, Patricia McGuire, Ludovic Phalippou, Geert Rouwenhorst, Moritz Schularick, Peter Scott, Jonathan Smith, Johannes Stroebel, John Tweddle, Stijn Van Nieuwerburgh, Joachim Voth, and Guillaume Vuillemey and two anonymous referees, as well as seminar and conference participants at HEC Paris, the LBS Summer Finance Symposium 2019, London School of Economics, the Luxembourg Asset Management Summit 2019, the SFS Cavalcade Asia-Pacific 2019, the University of Bonn Conference on Housing in the 21st Century, University of Geneva, and the World Economic History Congress 2018 for valuable feedback and data. David Klingle provided excellent research assistance. We are grateful to the archivists of King’s College and Trinity College, Cambridge, and of Christ Church and New College, Oxford for advice and assistance with data collection, and to the Cambridge Endowment for Research in Finance, the Centre for Endowment Asset Management, the Isaac Newton Trust, St John’s College Cambridge, ANR/Investissements d’Avenir [LabEx Ecodec/ANR-11-LABX-0047], and the Geneva Institute for Wealth Management for financial support. Any errors are our own. Supplementary data can be found on The Review of Financial Studies web site.

Footnotes

References