Preferences of Institutional Investors in Commercial Real Estate

The main variable of interest is the value of investors’ real estate portfolio. This allows us to construct the investor size quantiles (ISQ) of both buyers and sellers as explained in Section “Methodology”. We use two datasets (RCA, and RCA CPPIs) to construct this value. The RCA CPPIs (or Real Capital Analytics Commercial Property Price Indexes in full) are asset price indexes based on a structural time series repeat sales model given in Van de Minne et al. (2019). The RCA CPPIs are location and sector (i.e. property types) specific. In total we observe 80 of such RCA CPPIs.

First, we compute the average transaction price and transaction date for every unique property in the data. Next, we take the index corresponding to the market (location and property type specific) of the property, and assess the value of the property in any given year between 2001 and 2018. This gives us a full panel of all (appraised) property values between 2001 and 2018.

We also know from the RCA data who purchased and sold the property at time of sale of all the individual properties. This allows to find who owned which property between what time period. For example, if investor A sold a property in 2005 to investor B, we know that investor A owned the property before 2005 and investor B owned said property starting in 2005. There are some complexities though. For one, in approximately 5.5% of the transactions we find multiple investors (mostly designated “joint venture”), but we do not know the shares of ownership. In that case we give equal share of ownership to all investors. In some cases we also find that—for example—investor A bought a property in 2001, but investor B sold it in 2011. In this case we split the ownership in the middle, and assume investor B bought the property in 2006. In less than 4% of the cases did we find a missing transaction in the middle.

Figure 1 gives the average real estate portfolio size between 2002 and 2018, together with the 10th and 90th percentile of the portfolio sizes (grey shaded area). Average portfolio sizes first went up, from $70 million to $120 million, subsequently crashed and then after 2010, bounced back up again. In 2018 the average US real estate portfolio sizes are larger compared to its previous (pre-GFC) peak, at over $160 million. Note that the data is skewed to the right, because there are a few very large companies that increase the mean. For example, the median is closer to $10-15 million during this period. This is also visible from Fig. 2 by looking at the lower 10th percentile.

Table 1 gives summary statistics of the porfolio sizes by investor type. The first column gives the investor type. RCA has over 20 categories of investor types. For readability, we group them in the remainder of this paper into 4 larger groups (which are given in bold in Table 1); delegated, public, non-investor and direct. This categorization was used in previous literature as well, see for example Ghent (2018). Delegated investors include pension funds, equity funds, investment managers and banks. Public investor types include the real estate investment trusts (REITs) and real estate operating companies (REOCs). REITs and REOCs are under more scrutiny and are more transparent compared to the other firm types. In addition, these public companies have long holding periods by statute (Mühlhofer 2019), i.e. they make their money by collecting rents, and not by buying and selling of real estate. The non-investor group includes investor types finance, corporate, government, and non-profit. Finally, we have a group called “direct” which includes developer/owner/operators as well as all remaining investor types, like high net worth individuals and sovereign wealth funds (SWFs).

The second column of Table 1 gives how many firms of the corresponding investor type we identify in our dataset. The third column gives the average size of the firm’s US real estate portfolio per investor type, our main variable of interest. The fourth column provides the average value of the individual properties held by each investor type, and the final column gives how many properties they own on average (again in 2018). Values in different years are available upon request.

Note that the developer/owner/operator (DOO) investor type dominates the RCA dataset, and especially the “direct” group. Indeed, according to Table 1 the total value of all the commercial (investable) real estate is $4.7T in 2018 based on the transactions in our sample. The DOO investors own 48% of that. A large proportion of those DOO investors adopt the Limited Liability Company (LLC) structure which is typical for private equity funds. Those specialized investors may include family offices and small developers that act as investors rather than speculative developers. Overall, we think that those investors adopt the mandate to invest in real estate for investment purposes and not for owner occupation. Therefore, we think that while DOO is a broad category, it actually encompasses investors that specialize in investing in direct real estate. Their assets would almost entirely consist of properties and their main focus would be real estate investment.

Our summary statistics show that, on average, SWFs have the biggest real estate portfolios. However, note that only 15 companies are designated SWFs in our data. Their average real estate portfolio is worth over $3B in 2018. They own 45 properties on average, which are worth almost $80M each on average. This also means that on average SWFs own the most expensive real estate of all investor types. Non-profit organisations and governments are the smallest real estate owners on average. Their real estate is also the cheapest on average (together with religious buyer types) and have the least amount of real estate in general. Their portfolios are close to $40M on average in 2018. REITs own on average the most properties, with 72 followed by the non traded REITs with an average of 54 properties per non-traded REIT. The DOO investors, even though large by summation, are relatively small per investor. The average property value they own is almost $20M, and they “only” hold 5 properties on average in 2018, far less than for example (traded and non-traded) REITs and equity funds.

In order to get a feeling of how accurate our estimates of the investor sizes are, we will now take a detailed look at a selection of investors in the data. Table 2 gives the top 10 biggest real estate owners according to our estimates 2018. Unsurprisingly, number one is Blackstone with $81 billion in (US) real estate. For comparison, according to their own reports, Blackstone had $120B in assets under management (AuM) in real estate world wide in 2018. TIAA is the pension fund with the largest real estate exposure according to our estimates, with $33B. We also find three REITs in our top 10, Simon, Vornado, and SL Green. Given that REITs focus their investments primarily on real estate, the firm’s enterprise value should come close to the value of the underlying real estate (as valued by the stock market). As an extra robustness we therefore compare the (log) enterprise value of REITs taken from SNL Financial with our estimates of the (log) real estate portfolio value for REITs that we were able to match in Fig. 3. We match the REITs both, on firm and on an year level.

The overlap between our value estimates and the reported enterprise values is not perfect (Fig. 3), which might be caused by either stock market noise or inaccuracies in the RCA data, but overall there is a clear linear relationship between the two variables. It should be noted that our estimate of the real estate portfolio value is 10% less compared to the reported enterprise values of the REITs on average. This could indicate that REITs are overvalued (Van Nieuwerburgh 2019) or the other way around. However there are other explanations as well. First of all, REITs are mandated to invest most of its assets in real estate, but not all. More specifically, 75% of a REITs assets and income should come from real estate. Secondly, we know that private real estate returns typically lag public real estate returns. Reasons for this include informational inefficiencies and the slow (double-sided) search / negotiations in private real estate markets. As a result, previous literature found a lag of somewhere between 6 months to 1,5 years (Barkham and Geltner 1995; Francke and Van de Minne 2020a). Given that our analyzed time period has mostly been characterized by increasing prices, it should result in REITs having higher enterprise values compared to their contemporaneous private market valuations on average.

Next we will discuss the transaction data, which is relevant for our analysis. Note that in Section “Measuring Investor Size” we appraised a full panel of the firm values. In contrast, the transaction data is a very unbalanced panel of property level transactions. Properties are only sold once every 7–8 years. Table 3 gives the summary statistics of our most important variables.

Note that the average investor size quantile (ISQ, both for buyers and sellers) is 46 (the mean is between 1 and 92). This indicates that larger investors transact more on average. The variables which we believe are important in order to predict preferences/segmentation are: (1) closeness to the Central Business District, (2) size of the property (in square feet), (3) the quality of the structure (measured by a “Q score”, see below), (4) age of the structure, (5) whether the property is green certified, (6) property type, (7) Net Operating Income of the property at time of sale, (8) market, and (9) year of sale.

We get the NOI on the date of transaction from two sources. First, for about 35% of all transactions we observe the property’s NOI directly from the RCA transaction dataset. However, we also use mortgage data, also tracked by RCA. In many cases, whenever a property gets refinanced, an “underwritten NOI” is reported. Typically, there are a few years between the latest refinancing and the sale. We therefore change the NOI using “NOI indexes” from RCA. The NOI indexes are market (i.e. location and property type) specific. In total we observe approximately 300 of such NOI indexes in the US. This allows us to map an NOI at transaction for 48% of the entire dataset. NOI is important, because extant literature (Geltner et al.2014, p. 554–556) shows idiosyncratic price movement specific to individual assets or granular market segments, largely reflects the condition of rental (space) markets, as proxied by the NOI. At the same time, asset-valuation risk reflects changes over time in the capital market that cause changes in the opportunity cost of capital. Hence, time variation in the discount rate causes at least as much volatility in prices (see for example Geltner and Mei1995).⁴ However, the opportunity cost of capital is highly correlated over time across space markets. Mostly because the risk-free rate, which drives the cost of capital to a large extent, is identical no matter where the investment is made. The other determinant of the opportunity cost of capital—the risk premium—can differ per market, or even property, but rarely changes over time (Geltner and Mei 1995; Geltner et al. 2014). In further analysis we therefore also use time and location dummies to account for such offsets.

We find that approximately 6% of all properties in our transaction dataset are green certified. The apartment property type was transacted the most with 38%, and “only” 12% are industrial properties. In the appendix we also provide the distributions over year of sale (Table 9), location (Table 10) and investor type (Table 11). Given the often found positive price-volume correlation in real estate (van Dijk et al. 2019; De Wit et al. 2013), it is not surprising to find that the smallest percentage of transactions was during the GFC, in 2008–2010 (Table 9). Transaction volume is now higher compared to the previous, pre-GFC peak in 2007. Real Capital Analytics defines 134 different metro areas, see Table 10. Los Angeles is by far the largest market volume-wise, with 9.5% of all transactions being in said metro area. New York is second with over 6% of all transactions. In Table 11 we split the investors by buyer and seller. The “direct” investor group is again (see Table 1) the largest, being responsible for more than half of all the transactions.⁵ The “non-investor” group is the smallest investor type in our data. Finally, we find that approximately 10% of all transactions are done by a foreign investor.

Since the variation of the buyer size is our main variable of interest, we will explore this variable in more detail. First of all, if indeed specific properties are only bought by specific sized investors, we should find more variation in investor size in the data as a whole, than within the individual properties. To confirm this, we first measure the standard deviation of the ISQ within each property ID (if they are repeated observations). We subsequently order the standard deviations per property from lowest to highest. Some percentiles based on this ordering can be found in the second column of Table 4. We also subtract the average standard deviation of the ISQ of the entire dataset (26.7, see Table 3) from these percentiles in the third column. For completeness, we redo this exercise with the within investor ID standard deviation of the ISQ variable. Both property ID and investor ID will also be added as random effects in the frailty models.

Table 4 confirms that properties tend to be sold to similar sized ISQ investors. Only at the 83rd percentile is the within property σ^ISQ larger compared to the data as a whole. Within the investor IDs the distribution is even less equal. Only at the 99nd percentile is the within investor σ^ISQ larger compared to the data as a whole. In other words, once an investor is designated large, it will remain relatively large, and vice verse. Finally, note the final row of Table 4. Here we give the unique amount of properties and investors. We have 44 thousand observations in total and 30 thousand unique properties. Thus, most transactions are single sales. On average every investor was involved with over 4 transactions in the data.

Here we discuss the results of our baseline hazard model which can be interpreted as an ordered multinomial logit model looking at the size of the investors. Our baseline is the yearly de-meaned quantile distribution of size of the buyers’ real estate portfolio (ISQ_buyer). As discussed in previous sections, this is a variable which we constructed using the RCA transaction data and is unique in its nature. The explanatory variables are object characteristics, like size of the property and NOI, but also seller types. We have four different specifications. In the first (I) model we do not include any variables on the sellers. The second (II) model includes the investor size quantile, and model (III) introduces the seller types. We also include a dummy on whether the seller is a foreign investor or not. The ISQ score enters in the model with a log transformation. In the fourth (IV) model we enter the ISQ-score non-barometrically, i.e. we take up a separate dummy for every individual ISQ (\(q = 1, 2, {\dots } , 92\)) score. The estimates are expected to be somewhat noisy, however we have enough degrees of freedom to make inferences from the results. All models have year and metro level fixed effects. The results are given in Table 5.

It should be noted that the interpretation of these results can be less than straightforward, other than the sign and significance. A negative parameter estimates, means that larger (ISQ) investors are more likely to purchase the property, and the other way around. For example, the parameter estimate on the (log) size of the structure is negative and statistically significant in all model specifications, meaning that the larger the property, the larger the buyer on average. In contrast, older properties are more likely to be bought by smaller investors, all things equal as can be seen by the positive and significant effect of the log of the age of the property. Most parameters have the expected sign. Green properties, high Q score properties, and high NOI per square foot properties are more likely to be bought by largest investors. The distance to the closest CBD is mostly insignificant, or barely significant. This might seem surprising, but note that we control for metro fixed effects and property level NOI. Especially the latter is expected to be somewhat colinear with distance to CBD (i.e. properties further away from the CBD will have lower NOI per square foot). Apartment (industrial) properties are sold to the smallest (largest) investors, all things equal. Finally, even after controlling for the type of seller, we find that the size of the seller (ISQ_seller) has a significant impact on the buyer size. In all models we find that the larger sellers, tend to sell to larger buyers.

Again, even though the signs of the parameters are interesting in itself, it is difficult to get an intuition on the actual magnitude of the impact of said variables. We will therefore visualize the change in baseline (i.e. the probability distribution function (pdf)) for a specific set of linear combinations of covatiates. More specifically, we can make a “representative” property (Diewert et al. 2011; de Haan and Diewert 2011). This “average” property can serve as an overall benchmark. It has the average characteristics of the sample of all properties in the RCA database. Its values can be found in Table 3. This property is a mix of 38% apartment, 25% office, 25% retail and 11% industrial, it is 200,000 square feet, 28 years old, has an NOI of $12 thousand per square foot, and is 34 km away from the CBD. We subsequently run said covariates through our model estimates to get a pdf of buyers of that representative property. The pdf of buyers for Models (I) through (IV) is given in Fig. 4.⁶

Even though the models include very different covatiates, on average the pdfs are indistinguishable from each other. Indeed, in all cases small investors have relative small probability of purchasing this “average” benchmark property, but the same goes for the very large investors to the right of the pdf. The investor quantile most likely to purchase this average property has an ISQ score around the 50 mark for all models.

In the remainder of this section we take this benchmark property, but change a few property characteristics to see how much the pdf changes/skews. For example, we can take a few existing properties within the RCA dataset, and view the estimated probability to be sold to any investor size. We also include an online Appendix to this paper where one can simulate how changes to the covariates impact on the distribution of investors based on the size. Table 6 gives three examples of actual transactions in the RCA dataset. All estimates are based on Model (I), as for now we are not yet interested in sellers’ characteristics.

The first property is an old office property in the suburbs of Los Angeles. The property is up for redevelopment, and has no NOI as of the transaction date. Its a relatively small property (51 thousand square feet) and has the lowest Q score possible. Our model overwhelmingly predicts that a small investor would be interested in this property. The actual buyer was of the lowest category, i.e. the ISQ_buyer = 1. The middle panel is an apartment complex in Phoenix. The property has average scores throughout, and was bought (in 2006) by a mid-tier investor with an ISQ of 47. Our model also predicts that investors with a median portfolio size have the largest probability of purchasing this property. Finally, the third panel gives an industrial property, with a relatively low NOI per square foot. The property is younger than average, and has a Q score that is slightly above average with 1.71. However, our model predicts that only large investors are interested in this property. This is mostly caused by the fact that this is one of the largest properties in the dataset with 1.2M square feet. Above results can be explained in several ways. One the one hand, the findings are indicative of potential financing constraints as the small-sized buildings are predominantly bought by small investors and the large buildings by large investors. Small investors may face financial constraints to buy large and hence more expensive⁷ buildings. On the other hand, investors may prefer to buy certain buildings as they choose to adopt different investment styles. For example, small investors may adopt a value-add or opportunistic strategy by buying properties with low Q score and low NOI, then refurbish those and benefit from the capital gains. Large investors instead may be interested in the so called ‘core’ buildings which tend to be large and more expensive. They may be more interested in stable cash flows instead.

A useful outcome of above exercise is that we are able to pin down what a core, a value-add and an opportunistic building may look like if we want to keep working with the common terminology. The property characteristics associated with core buildings may vary depending on the sorting of investors across properties. Looking at which properties get bought by the top end of the pdf—the large players—can help us identify what core is for those investors. One example is the real estate displayed at the bottom panel in Table 6. In our example, a very large young industrial building with high quality and high overall NOI. A value-add property may look like the real estate displayed in the middle panel. And an opportunistic—at the top panel. The latter is a small old office building with very low NOI and the lowest possible quality. It is still located in a major city—Los Angeles—but it is away from the CBD.

In Fig. 5 we keep the average representative property, as given in Fig. 4, however we change two characteristics: (1) NOI per square foot, (2) and square footage itself. More specifically, we look how investors sort themselves across properties varying the combination of low/high NOI and small/large structures. The combination of low and small is designated to be at the 15th-percentile of investor size. That is associated with $5.09 NOI per square foot, and 39 thousand square feet. That means that small investors would buy the small properties with low NOI per square foot all else equal. The high/large property is designated to be bought by the 75th-percentile ISQ. That us associated with an NOI per square foot of $16.26 and 258 thousand square feet. From Fig. 5 we can clearly see that the largest “jump” comes from the building size. Even at a low NOI per square foot, larger investors still prefer larger properties. It does drop off a little for the very largest investors, although it should be noted that the total amount of NOI (NOI per square foot times the square footage) is twice as high for the combination of low NOI and large building ($1.3M), compared to the high NOI and small building combination ($0.6M). That means that large buildings have larger NOIs per square foot than smaller buildings.

Next we are interested in some of the time-varying effects associated with buyer preferences. For this we use our benchmark property again but we change two variables. The first is the NOI per square foot. This is a variable that is affected by the economic environment of that year. We use the means that are shown in Fig. 2. The second variable we change are the year dummies, to correspond with the year of interest. By doing this, we can see if the preference for the same property changes over time, by holding the other characteristics of said property fixed. We are specifically interested in the GFC period which spands from 2007 to 2009, for reasons we will show below. Because we are still not interested in the seller characteristics, we show the results of Model (I), although other models remain very similar and are available upon request. The results for a selection of years are given in Fig. 6.⁸

What is interesting about Fig. 6 is that most of the changes of the distribution of buyer ISQs during the GFC is only affecting the higher ISQ investors. Note that the pdfs always sum to one. Thus, one cannot say that the probability of sale changed from one year to the next, as all results are to be interpreted within the confines of the given year. That being said, it is evident that the larger investors “pulled out” of the real estate market, compared to other investor sizes. Instead, the probability that the average property was bought by lower and mid-tier investors increased. The opposite picture is observed just prior to the crisis, with the largest investors being the most likely to purchase the average property. In 2009, the probability that a large investor would purchase the average property was at a historic low. (Since then, the probability has gone up again, but not as high as the immediate pre-GFC levels.)

Above observations are somewhat counter-intuitive as one would expect that the large investors have the least credit constraints and in a crisis period would be the predominant player on the markets. Clearly, they were the ones that actually have reduced the likelihood of buying the benchmark property. That finding suggests that they seem to be the most responsive ones to an economic downturn by deciding not to trade on the real estate market. An economic downturn may therefore be associated with the largest investors more strongly lowering trading activity than small and medium investors. This can be because large players wait for the real estate market to recover before they start buying or selling. They may be the least flexible to manoeuvre the GFC market and find good deals. Alternatively, those investors may struggle to justify a transaction during the GFC for their shareholders or lenders. This is because large investors (1) are subject to more stringent regulatory requirements, (2) may be listed on the stock exchange and subject to more scrutiny from shareholders, (3) or/and have strong dependence on issuing bonds and hence be wary of their credit ratings. Pension funds and insurance companies may have been affected by the introduction of new financial market regulations in the US and in Europe (i.e. Dott Frank, Basel III, Solvency II, etc.). REITs may face devaluation of their assets and as a result may find it harder to refinance and prefer not to opt for new stock issuance. Independently from which channel dominates, the conclusion is that being a large investor and potentially having established good and cheap financing structures does not result in increased or even a continued trading activity in a downturn. This may be associated with a shock to liquidity for those large players. Smaller players may face financing constraints and illiquidity too but it seems they are more agile during a crisis and keep trading activity going. That can be linked to style investing—small investors look for opportunistic properties and the possibility for such deals increases in a downturn. The fact that the small players continue to trade in the GFC suggests against the presence of a financing constraint channel that explains the pdf of ISQs.

Above findings suggest that the potential effects of a change of the market composition of buyers and sellers may lead to overall lower prices in addition to any other downward pressure on prices during the GFC. That means that the presence of large investors and their diminishing likelihood to transact during a downturn and their higher likelihood to transact in a growing market can exaggerate the real estate cycle. The presence of small players may have the opposite effect on the real estate market and can help smooth the cycle.

This is also buttressed by looking at the expected buyer sizes per state for a selection of years, see Fig. 7. The expected buyer size is the sum product between the probabilities and the ISQ score of the pdfs. To construct Fig. 7 we first construct an average benchmark property per state. The average structure size, property type distribution, etc., is thus different for every state. Next, we change the average NOI per square foot per state per year, and the year of sale dummy. These estimates are still based on Model (I). States with less than 10 observations per year are omitted, and are given in grey.

We find that in 2017, the distribution of expected buyer sizes across states is approximately similar to that in 2003, albeit a bit less concentrated in New York and Chicago (Illinois). In fact, the largest investors in 2017 can be found in DC and Massachusetts (dominated by Boston metro). Just pre-GFC in 2007, we find that the average size of investors increased. It did so quite in-discriminatory. In other words, larger investors become more active in almost every market. Two of the exceptions are the states of Michigan and Utah (dominated by Detroit and Salt Lake City respectively) where the average investor size actually went down. More interestingly, the state of New York (dominated by New York City) also saw decreases in average investor sizes between 2003 and 2017. States that saw its investor sizes increase the most between 2003–2007 were Wyoming, Ohio, Kentucky, and Alabama. These are states that are not typically known to attract large institutional investors. In the build-up to the GFC hence there seems to have been a shift of large players moving away from the more established commercial real estate markets into states with previously less institutional grade properties. That may have been due to a search for yield and better investment opportunities. At the end of the GFC, large institutional investors reduced their real estate investments in most states. In fact, none of the states had an average increase in institutional investor sizes. The largest drop in investor size between 2007–2009 was in Connecticut, followed by Ohio. Large institutional investors found their way back to real estate in 2017, although not equally in every state. States that only saw a small increase (none of the states had a decrease between 2009–2017) were Wyoming, Indiana, Missouri. In levels, the smallest investors can be found in South Carolina in 2017.

Next, we are interested in the size of the sellers‘ US real estate portfolios (ISQ_seller). In models (II) and (III) we include the log of the ISQ of the seller. In model (IV) we use a dummy for every of the 92 ISQ categories of the sellers instead. The estimates of the dummies are graphically given in Fig. 8. Note again in Table 5 that the estimates on the covariates hardly change between the models, so earlier results still hold for these models as well. Here we are mostly going to discuss the added variables on the sellers.

For models (II) and (III) we both find a negative effect of the log of the sellers’ ISQ on the buyers’ ISQ, meaning larger sellers sell to larger buyers, holding all things equal. Model (IV) gives a similar image. Even though the estimates can be a bit “noisy” at times, overall it is clear that the larger the sellers’ ISQ, the larger the buyers’ ISQ. If there would be no preference to trade with a specific size of investor, these dummies should be insignificantly different from zero. The results are very robust.

Next we compare the effect the sellers’ size has on the probability to sell it to a buyer of a certain ISQ. For this we again use the benchmark property. The only parameters we change are the sellers’ ISQ. More specifically, we are going to compare the buyers ISQ pdf when the sellers’ ISQ is 15 versus 75. We do this for all the models (II through IV) that include the ISQ_seller variable.

Even though the parametrization is very different between the models (two include a log transformed ISQ score, and in the other ISQ is entered non-parametrically), the results are very similar. The results are presented graphically in Fig. 9. In all cases we find that if the seller has an ISQ of 15, i.e. the seller is small, it is likely to sell to a smaller investor on average. The average ISQ_buyer is expected to be 44. This is compared to when the seller is large, i.e. they have an ISQ of 75, in which case the average expected ISQ_buyer is approximately 50 on average. Again note that it is the same property in both cases, and the same investor type. The main difference is—similar to our GFC analysis in Fig. 6—in the high ISQ buyers. The probability that a smaller investor will sell the same property to a large investor is relatively slim. Large investors overwhelmingly sell to other large investors, irrespective of the type of investor. These findings align with what has currently been observed in the industry where investors within the same size and background sell to each other. Above findings suggest that the real estate market is segmented and investors trade with each other in separate brackets depending on their size. That may be due to established networks within a given investor sub-market and familiarity with the players. Large sellers are mostly selling the average property to large buyers while small sellers can sell to a range of buyers but the likelihood to sell to the biggest buyers is very small. It is important to keep in mind that in each of those scenarios, we are dealing with the same property and not with larger or smaller properties depending on the size of the seller. That may suggest that the large sellers mostly sell properties within their network and that the network effect is stronger for the very large investors.

Finally, we will shortly discuss the model diagnostics. The differences in the concordance score (Harrell et al. 1996) between models (I) through (IV) are negligible although the likelihood and AIC do improve after introducing sellers characteristics.⁹ The AIC prefers the third model, but the likelihood the fourth model. The Wald test shows that all parameter estimates are different from their starting value.

Next we will discuss two sets of auxiliary results. First we will discuss our hazard models with frailty for individual properties and investors. These results can be found in Table 7. In Model (V) we only include the property level random effects, in Model (VI) only the investor level random effect, and Model (VI) contains both random effects.

Only including the random effects for the properties does not alter the results erratically. It seems in general that there is not much unobserved heterogeneity with this set of covariates, evident from the low standard deviation on the random effects (σ_{k=property id} = 0.006, see Table 7). Only the green certification and the dummy for direct investor seller type become insignificant. Including the investor random effects does attenuate most of the results. For example, none of the seller type variables are significant anymore. Also the effect of Q-score reduces, but is still highly significant. The coefficients on property types also reduce considerably. This can be explained by the fact that most investors tend to focus on one sector, thus the investor random effect effectively controls implicitly for property types. The same argument can be used to explain why the age variable becomes insignificant; Investors tend to focus on specific vintages. Interestingly, the effect of NOI per square foot and the size of the structure remain very robust. The effect of the prior investor size (ISQ_seller) stays highly significant, but is attenuated as well. All of this is very encouraging, given the low amount of variation we found within investors’ ISQ (Data and Descriptive Statistics of the Other Variables) and the fact that the standard deviation on the investors random effects is quite large (σ_{k=investor id} = 3.2 and 3.5, see Table 7) for both models with the investor random effects. Finally, the fit—according to the marginal log likelihood—is best for the model with both random effects included.

Next, we decide to swap out our dependent variable—ISQ—with a variable for how much prior exposure the investor had in a specific market before the transaction. We—again - use the 80 market definitions provided to us by RCA (which are location and sector specific), and compute the dollar value the investor already had in said market prior to the transaction, as a percentage of the total value of the investors’ real estate portfolio. As with the ISQ, we make quantiles of this variable by grouping investors who had less than 5% of their portfolio in a market and 95% or more in a market, and with 1% increments in between, resulting in—again—92 quantiles. We call this variable PEX. This variable is of interest to us, as it highly relates to the ISQ variable. The downside is that it is heavily endogenous, and we therefore cannot simply add it to our main specifications. Indeed, only larger investors have the capacity to diversify. For this reasons we estimate a separate model and see if our findings are consistent. The results are given in Table 8.

Note that the results of our PEX model mirrors the findings of our main models in Table 5. The reason being is that larger investors tend to have less prior exposure on average per market because they are relatively diversified. Thus a positive estimate on ISQ, translates to a negative estimate in the PEX model.

From our results, we find that investors with little prior exposure in a market tend to purchase properties that are larger, green certified, high Q score, closer to the CBD and of low age. These properties are arguably more informationally efficient (Geltner and van de Minne 2017), and tend to be “safer”/core investments. This could also explain partly why larger investors tend to go with these larger properties. Investors with a lot of prior exposure also tend to purchase apartments, whereas industrial and retail is more likely to be bought by investors with less prior exposure to the market. As was the case with the ISQ variables, we also find that the sellers’ prior exposure (PEX_seller) explains the prior exposure of the buyer. “Local” investors (high PEX) tend to sell to other “local” investors, and the other way around as documented by the significant coefficient of PEX_seller in Models (II)–(IV). The only estimate that is inconsistent with previous findings is that of NOI per square foot. This estimate is insignificant for all models, whereas it had a large impact on ISQ.

Even though we will not show all the pdfs you can make from the results in Table 8, we do want to highlight one of the results. Figure 10 gives the pdf of the “average” representative property being bought by an investor with little prior exposure for different years. Little experience is defined as having less than 5% of its portfolio in the target market prior to the transaction. We find that on average, our representative property has a 13%–14% change of being bought by an investor with little experience in that market. This number is relatively stable for the different years. However, during the GFC we find a few basis points drop in this probability. This could mean that during the crisis investors were more afraid to invest in markets they did not know well.

Finally, we present alternative estimations based on the multinomial logistic regression model and looking at investor types instead of size. Our dependent variable is categorical and consists of the four investor groups we constructed in Section “Data and Descriptive Statistics”. We use the same explanatory variables as in the baseline model (I). The actual estimation results, with model fit and significance can be found in the Appendix, see Table 12. The marginal effects of the estimated coefficients can be found in Table 13. Note that the rows always sum up to zero by construct.

Starting with the delegated investor, we find that they prefer large, high Q score, high NOI, green certified, young properties, close to the Central Business District (CBD). A negative coefficient for CBD means that the further away the property is from the closest CBD, the lower the probability a delegated investor will purchase it. In terms of the Q score, for every 1% increase in it, the probability that a delegated investor will purchase the property, increases by 0.15%. In other words, delegated investors are interested in “A-class” stabilized properties, essentially “cash-cows”. Delegated investors also prefer industrial properties the most as compared to the other investor types. Those industrial properties mostly consist of logistics centers and warehouses. Delegated investors are also the ones most likely to buy real estate from foreign sellers.

Public investors fit that profile to a large extent as well. Indeed, they also prefer large properties (positive coefficient on structure size, with 0.025), young properties and high Q scores. Compared to direct and the non-investor groups they also value high NOI per square foot more, but not as much as delegated investors. This effect is insignificant (Table 12). In contrast to delegated investors, public investors care less about proximity to CBDs. In general we find that public investors are spatially well diversified, as is to be expected. Relatively speaking, public investors buy less of buildings which have green certification. Also, they are most likely to buy retail properties.

Direct investors seem to have the complete opposite preferences compared to delegated or public investors. Compared to those investors, they buy smaller properties, with low Q scores, that are older, with relative low NOI per square foot. In other words, the direct investors are more opportunistic and seek to add value. They are most likely to purchase apartments. The non-investors are middle of the road in almost all categories. Compared to the other investor types they are not interested in particularly large or small, high or low Q score properties.

Finally, we will discuss the bottom panel of Table 13. Here we find the estimates for the seller type. The results indicate that—holding all things equal—direct investors are most likely to purchase from other direct investors; non-investors buy mostly from non-investors; public investors have the highest probability to buy their real estate from other public investors. Only delegated investors mostly buy from an other category of investors different to theirs, namely the public firms. However, their second favorite seller are other delegated investors.

In summary, we find that in the absence of investor size (ISQ), investor type has been used to sort investors across properties. Larger, high quality properties are mostly sold to delegated and public firms. This is in line with our findings about large investors.

Furthermore, we find that investors tend to trade within their own investor type, holding constant for the property quality. This means that even if another type of investor had a property with the same characteristics—investors would choose to buy from the same category as theirs. This may be due to network effects—investors within the same category belong to the same networks and may feel more accustomed to do business with familiar players. We will not test the channel at hand in this paper but it remains for future research to identify the mechanisms behind this sorting between buyers and sellers.