Venture Capital and the Market for Talent during Booms and Busts

1. Introduction

The venture capital market is strikingly cyclical—seemingly more so than is justified by changes in real investment opportunities. Feast-or-famine availability of funds leads at times to a money chasing deals environment (Gompers and Lerner, 2000) with correspondingly low returns. Yet there appears to be more at work than merely periodic overvaluations. Kaplan and Schoar (2005) argue that the intrinsic quality of general partners (GPs)—those who actually manage the funds—follow predictable patterns. During booms, new entrant venture capitals (VCs) are prone to excess while established, proven VCs remain relatively sober in their investment decisions. Few of these inept new entrants raise follow-on funds.

In this paper, I develop and test a model which naturally captures these phenomena and delivers new predictions. Entrepreneurial agents in the model either start new firms (with VC financing) or continue in their current employment. There is a continuum of such agents, only a fraction of whom enter the market in equilibrium. The supply side also has latent agents, that is, potential information producers who become VCs if market conditions warrant doing so. Agents on both sides vary in skill. Latent supply side agents generate, in effect, an aggregate supply curve of information production. Moving up the supply curve corresponds to picking out progressively less capable VCs. The model therefore has two layers of asymmetric information. With free entry on both sides of the market, supply and demand are both endogenous.

The model has the following properties. In all states, active agents are of higher quality than inactive agents, and too many projects receive funding. That is, there is overinvestment in the sense that the worst active entrepreneur has a negative NPV project. Given a positive shock to investment opportunities, new agents enter on both sides of the market but, importantly, these agents are of relatively low quality. Thus, there is a sense in which frictions are procyclical. This provides a theoretical structure for understanding Kaplan and Schoar’s (2005) finding that boom time finance deals are disproportionately handled by inept agents.

The model also makes several new predictions. First, during hot markets there is a wider dispersion of entrepreneurial quality. Consequently, as the quality of projects is ultimately revealed (e.g., after liquidation or IPO) this leads to heightened dispersion of returns across portfolio firms during booms. Second, because there is a wider dispersion in VC quality as well, a similar cross-sectional prediction holds at the VC fund level.

I test this second prediction using fund-level data from Preqin between 1982 and 2004. I classify the sample according to 1) whether the fund’s vintage year is hot or cold, 2) whether the parent VC firm is an incumbent or new entrant (i.e., whether this was their first fund), and 3) whether the parent firm was founded in a hot market or cold market. Dividing into these subsamples yields second moments consistent with the model’s predictions. For example, new entrants in hot vintages have much higher cross-sectional dispersion in abnormal returns than new entrants in cold vintages (39.2% versus 15.8%; p-value  < 0.01). For incumbents, hot vintages are also associated with somewhat higher dispersion, but the effects are more modest.

Closer examination of the full distribution of abnormal returns shows a pronounced difference in the left-tails of these subsamples. For incumbents, the 10th percentile of abnormal returns is −12.9% and −18.5%, in cold and hot markets, respectively. For new entrants in cold markets, the 10th percentile is −14.6%, only slightly trailing the incumbents. For new entrants in hot markets, the 10th percentile of abnormal returns falls to fully −27.6%.

Preqin also categorizes fund returns into performance quartiles Q1 (best) through Q4 (worst). The precise benchmarking is proprietary and is based on a variety of factors, including vintage year, industry focus, and geographical region. Consistent with the model, new entrants in hot markets are disproportionately likely to be assigned the worst quartile ranking Q4 (30.1%) whereas incumbents in hot markets have a relatively low chance of this event (20.6%). In fact, the incumbent’s distribution is nearly invariant to market conditions. In contrast, new entrants in cold markets are unusually likely to earn the top performance rating (28.5%). This finding is consistent with the model’s prediction that cold markets tend to involve high-quality agents.

More broadly, the model’s relevance extends beyond venture capital. Oversupply of low-quality deals will eventually spill over into other financial markets. As these poorly run firms struggle and stagnate, opportunities arise for private equity or event-driven hedge funds to invest in distressed securities, including restructuring, recapitalization, or bankruptcy-related asset sales. These alternative financing markets likewise need to attract new talent to handle the increased demand. A similar story will play out as these sectors become overcrowded—ultimately, with low-quality entrants. Thus, cycles in one financial market propagate waves and then crashes elsewhere in the financial ecosystem. The model in this paper provides a starting point for understanding how these cycles are related to the intertemporal distribution of talent across industries.

1.1 Related Literature with Sticky Short-Run Supply

It has been argued that the supply of venture capital funds is “sticky”.¹ Private equity investing requires considerable skill and experience—attributes which are largely fixed in the short-run. During a hot market when demand shifts out, established GPs become a scarce resource and therefore earn rents. As emphasized by recent theoretical literature, these temporary supply-demand imbalances alter the relative bargaining power of VCs and entrepreneurs. See Inderst and Muller (2004); Kanniainen and Keuschnigg (2003); Khanna, Noe, and Sonti (2008); and Koskinen, Rebello, and Wang (2008) for analysis of how these imbalances affect effort choices, portfolio size, banking wages, and security design, respectively.

The present model, by design, does not assume fixed supply. This omission is not intended to reflect the author’s stance on the empirical speed of adjustment of supply and demand. Rather, the model can be viewed as describing the equilibrium distribution of talent once supply and demand fully adjust to investment opportunity shocks.

1.2 Related Literature with Endogenous Demand

Khanna, Noe, and Sonti (2008) and Yung, Colak, and Wang (2008) develop theoretical models in which hot markets are associated with lower quality issuers on average—a result which is echoed here. Their modeling of the supply side is very different from that of the current paper, however. In Khanna etal., the set of supply-side agents is fixed by assumption, and these agents are identical in skill. Underwriters are viewed as intermediaries who hire these skilled information production agents. The key step in the analysis lies in endogenizing these agents’ wages. With fixed supply, wages rise in a hot market—capacity constraints become more important—and so underwriters economize by reducing their demand for screening labor. In Yung etal., the supply side is effectively unmodeled: there is unlimited capacity and no screening. The current model generalizes these approaches to an environment with free entry of heterogeneous agents on both the demand-side and the supply-side.

1.3 Related Literature with Double-Sided Moral Hazard

Because both the entrepreneur and the VC have real inputs to firm value, it is natural to model the environment as one of double-sided moral hazard. This paradigm has important security design implications (Repullo and Suarez, 1999; Casamatta, 2003; Schmidt, 2003).²

By focusing on how these frictions vary with business conditions, two papers in this literature are particularly close to the spirit of the current model. Inderst and Muller (2004) model the process of search and bargaining between VCs and entrepreneurs. The relative supply and demand for funds determines the equilibrium stake held by VCs and the residual claim held by entrepreneurs, and in turn, the amount of effort chosen by each party. Total firm value can be nonmonotonic in the amount of capital market competition because at low (high) values of competition for funds the VC (entrepreneur) contributes too little effort. In Inderst and Muller’s analysis, shocks to investment conditions have a dynamic effect because the supply of venture capital is assumed to be sticky in the short-run. Therefore, given a positive shock, VCs temporarily earn positive profits. In the long-run, these profits are driven away by the entry of new VCs. (The severity of frictions can either go up or down, depending on the pre-shock sharing rule.)

Kanniainen and Keuschnigg (2003) take an alternative approach to modeling the effect of changing business conditions. They demonstrate that when industry returns are high, VCs spread their labor thinner by holding a greater number of portfolio firms and hence contributing less effort to each firm. Hence, as in the current model, there is a sense in which the average supply-side “quality” deteriorates during good times. However, in that model VCs are homogeneous and all take the same action. In contrast, the defining feature of the current model is that the distribution of venture capitalist quality changes over time. This heterogeneity leads to time-varying dispersion in outcomes.

2. The Demand for VC Finance

There exists a continuum of agents in the economy with nontransferable access to an entrepreneurial project. Undertaking the project requires that the entrepreneur forgo labor income worth V.

In this section, to provide a benchmark against which Section 3 will be compared, I assume there exist many identical agents who can become venture capitalists if doing so is profitable. Payoffs to the project are 1 with probability π_i and zero otherwise. Entrepreneurial projects require K units of capital to function, otherwise the payoff is zero for sure.³ The subscript i indicates the entrepreneur’s privately known type.

Venture capitalists also serve an information production role. They can evaluate entrepreneurs, obtaining a signal s  =  G or s = B. This signal is associated with two costs. The VC incurs an information production cost C_{V C}, and the entrepreneur pays a nonpecuniary effort cost C_E (interpreted as the opportunity cost of meeting with VCs, preparing the pitch, etc.).

The relevant distinctions between the aforementioned costs are the following. While the entrepreneur bears C_E directly, he bears C_{V C} and K only indirectly through its effect on the venture capitalist’s required stake. Finally, the opportunity cost V is borne only if the project is undertaken.

Condition (1) implies that the signal is informative: higher quality firms are more likely to yield the signal s = G. To rule out the scenario in which purchase of these signals is socially suboptimal—the “uninformed finance” environment is not of interest here—assume a mass of entrepreneurs with $πMIN=0$ who pool if the VC operates as an uninformed agent in equilibrium. For example, one could assume that they have zero opportunity costs and a small but positive private benefit to running the firm.⁴ This assumption captures the spirit of Kaplan and Schoar’s (2005) intuition that if VC did no due diligence, the candidate pool would be flooded with bad projects. Given assumption (1), however, such entrepreneurs can never secure funding: they always obtain the signal s = B. Accordingly, these entrepreneurs do not affect the comparative statics of the equilibria; rather, their presence is used to rule out “uninformed” finance scenarios.

The sum of the aforementioned opportunity costs must be less than one (i.e., $K+V+CE+CVC<1$⁠) otherwise no entrepreneurs are worth funding. Assume that success probabilities π_i are uniformly distributed on $[0,1]$⁠. Hence, some entrepreneurs have positive NPV projects while others have negative NPV projects.

Some fraction of the potential entrepreneurs opt to remain in the labor pool rather than enter the VC market. Let π_MIN denote the lowest quality entrepreneur that tries to obtain VC financing. This cutoff value will be determined later. The interval $[πMIN,1]$ then determines the number of active entrepreneurs. It is assumed that entrepreneurs are evenly distributed across VCs so that each VC has n active entrepreneurs. Evenly distributing entrepreneurs across VCs in this way ignores an “integer” problem, that is, the number of entrepreneurs is unlikely to be divisible by the number of VCs, so that each pool cannot literally be of exactly the same size.

Naturally, because the signal s is informative, $E{π|s=G}>π¯$⁠. Both the unconditional mean and the conditional mean are increasing in π_MIN. An exogenous improvement in the entrepreneurial pool implies a better expected quality conditional on passing the venture capitalist’s screen.

2.1 The VC’s Problem

2.2 The Entrepreneur’s Problem

The bracketed terms indicate the expected payoff if the project is funded or unfunded, respectively, while the leading terms reflect the probability of each outcome. Note that because entrepreneurs know their type, from their point of view the venture capitalist’s signal adds no value-relevant information. Rather, the signal only determines whether the funding proposal is accepted.

Putting these two entry conditions together yields the following result. 

The inequality (7) is an incentive compatibility condition. In particular, Equations (4)–(6) are derived assuming that the signal is purchased, whereas Equation (7) guarantees that this purchase is indeed optimal at the equilibrium. Recall that if instead the signal is not purchased, then the market is flooded with bad entrepreneurs (in which case, again, purchasing the signal becomes optimal) and thus the conditions in Theorem 1 are both necessary and sufficient.

Non-existence can be motivated heuristically as follows. The left-hand side of Equation (7) is negative when π_MIN and C_{V C} are high. But when $πMIN≈1$ there is virtually no dispersion in the pool of active entrepreneurs. This lack of uncertainty negates any incentive to purchase the signal in the first place: VCs would unilaterally defect to non-purchase, which violates the assumptions employed in deriving {(4), (6)}.

To summarize, the solution to {(4), (6)} must involve a nontrivial amount of heterogeneity in entrepreneurs in order for VC screening to be incentive compatible. On the other hand, the entire thought exercise contained in this paper is of little interest without significant dispersion in agent quality.

Before turning to the model’s comparative statics, recall that the payoffs of a successful firm are held constant at one. Therefore, decreasing either K or V (relative to one) serves to increase the NPV of the firm’s project and should be considered therefore as a positive shock to investment opportunities. It may seem more natural to model such positive shocks by simply increasing the terminal value—and indeed an earlier draft of this paper did so. However, doing so is mathematically equivalent to simultaneously reducing the cost of all inputs relative to terminal output. This alternative approach has an important drawback: varying the cost of each input relative to terminal value has different effects, as shown below. These effects would be obscured by considering only a uniform shock to the terminal value. 

This outcome has the comparative static properties motivated in the introduction. Equations (8)–(11) indicate that as investment opportunities improve, lower quality entrepreneurs are drawn into the market. That is, as V and K shrink (relative to terminal payoffs of a successful project), the cutoff quality π_MIN drops. In this sense the demand-side asymmetric information problem is of procyclical severity.

Equation (8)–(11) also indicate how VC contracts respond to changing market conditions. Unsurprisingly, as the required VC effort costs rise, VCs demand a larger stake. The result $∂α∂V<0$ has more subtle intuition. An increase in the value of outside opportunities (relative to VC finance) leads to improvement in the quality of the entrepreneurial pool. Because of this reduction in the severity of the asymmetric information problem, VCs are satisfied with a smaller stake. Putting these two results together, the model is ambiguous regarding how α varies with the value of investment opportunities. If both V and K shrink relative to terminal payoffs, then there are competing effects on α; this issue is explored again in a numerical example in Section 3.3.

Any change which increases a VC’s costs causes α to rise. Any changed which increases the entrepreneur’s costs causes α to fall, since it improves the quality of the pool, allowing the VC to demand a smaller stake.

3. The Supply of VC Finance

In Section 2, venture capitalists are identical and perfectly competitive, which drives their profits to zero. This modeling assumption precludes analysis of incentives to enter the VC market. To address this issue, Section 3 introduces heterogeneity in VC skill.

3.1 Two Venture Capitalist Types

Assume that some VCs have information production cost C_L and all others have cost C_H>C_L. Importantly, both types of venture capitalists obtain the same signal; it is only the costs of production that differ. Consequently, there is no post-financing difference between VCs and so, for a given contract, entrepreneurs have no strict preference regarding which VC supplies funding.

Further assume that there are some states in which the units of VC labor demanded exceeds the endowment of low-cost VCs. That is, low-cost VCs are in short supply, and this is a binding constraint. Thus, high-cost VCs must also be active to ensure that markets clear.⁷

This equilibrium has the following properties. High-cost VCs break even. Competition between them cannot reduce $α*$ without violating the equality in Equation (13). Low-cost VCs earn expected profits $CH−CL>0$ from their evaluations.

3.2 Continuum of VC Types

Section 2 demonstrates that venture capital labor demanded is a decreasing function of the cutoff quality π_MIN. When π_MIN is high, few entrepreneurs are active and therefore labor demand is low. More generally, denote this demand function $D=D(πMIN)$_, where $D′<0$⁠.⁸

Next suppose that each potential VC is associated with an information production cost C_i drawn from some continuous distribution. Following the logic of Section 3.1, there is a cutoff C_MAX indicating the highest-cost VC active in equilibrium. The venture capitalist with $Ci=CMAX$ earns zero profit. All those with cost exceeding C_MAX drop out of the market, while those with cost less than C_MAX earn strictly positive profits from evaluating entrepreneurs. Therefore, in general, the measure of active agents generates a supply of venture capital finance $S=S(CMAX)$ where $S′>0$⁠.

The demand equation comes from substituting $πi=πMIN$ in Expression (5). The left side of the equation is the profit of the worst entrepreneur active in equilibrium. By definition, π_MIN is the entrepreneurial quality associated with zero profit. The supply equation comes from substituting C_VC=C_MAX in Equation (4). That is, C_MAX is the VC cost associated with zero profit. Finally, markets must clear, that is, the aggregate supply must equal the aggregate demand. 

The existence condition is analogous to that following Theorem 1. It ensures that the marginal VC has an incentive to purchase the signal rather than supplying funding on an uninformed basis. Specifically, she earns zero profit as an “informed” agent but would earn negative profit as an uninformed agent. Precisely the same condition ensures that sidelined VCs would earn negative profits were they to enter as an uninformed source of finance. 

The inequalities in (15)–(17) formalize the intuition laid out in the introduction, examining how the double-sided asymmetric information problem depends upon real investment opportunities. A positive shock to the economy (as indicated by a drop in K) has several implications. Most obviously, venture capitalists’ required stakes fall because they are being asked to provide less costly effort while terminal payoffs are unchanged.

A key property of this asymmetric information paradigm is that π_MIN drops in response to a positive shock to investment opportunities. Financing frictions worsen because as the set of real opportunities improves, low-quality entrepreneurs enter the market. Analogously on the supply side, C_MAX deteriorates in response to this economic shock. Thus, hot markets induce the entry of low-quality agents on both sides of the market.

The model also makes a related observation regarding the rents to high-quality information producers. In Section 3.1, it was determined that the profit to high-quality VCs is zero when they are the only agents active (i.e., states of low demand) but their profit is $CH−CL>0$ in states of high demand. That is, high-quality agents’ profits depend upon the difference between their own quality and that of the worst active agent. Analogously here, since the quality of the worst agent is continuously decreasing in market heat, the profits of a fixed (active) VC are continuously increasing in market heat. As a result, total profits to the VC industry are also increasing in market heat.

Equations (16) and (17) have straightforward interpretations. As the relative costs to entrepreneurs of seeking VC finance increase, marginal entrepreneurs are forced out of the market—that is, π_MIN grows. Note that even though VCs do not bear the entrepreneurs’ opportunity costs, there is a channel of indirect effects. Specifically, as these opportunity costs rise, weak entrepreneurs drop out. As the quality of pool improves, this enables VCs to demand smaller stakes.

The final indirect effect is that because demand falls, the equilibrium number of VCs operating must fall. Such a drop would be impossible if α rose, because in that case VCs would make higher profits even for a fixed level of asymmetric information. These higher profits would encourage entry, not exit.

3.3 A Numerical Example

Consider ${CE=0,V=.2,K=.4}$⁠. Assume VC quality has a discrete distribution, with some VCs at C_L = 0. Initially suppose that these low-cost VCs are sufficiently numerous to satisfy demand. Next, consider a shock to ${CE=0,V=.19,K=.39}$⁠, representing a small improvement to investment opportunities. Suppose that the zero-cost VCs can no longer satisfy demand, necessitating another batch of VCs with higher cost (say $CH=.01$⁠) to enter the market.

This outcome exemplifies the theme of the paper. An improvement in investment opportunities induces new agents to enter, and these agents are of relatively low-quality; that is, π_min drops.

Thus, risk increases in hot markets.

The first term in each decomposition is variance coming from portfolio firm heterogeneity. The second term is the variance introduced by VC heterogeneity.⁹ Hot markets are associated with increased variance in both components. To summarize, VC fund returns have more dispersion during hot markets, because of increased dispersion in both firm quality and in VC quality.

4. Empirical Results

I now turn to testing the model. It is helpful to define a few terms used in data. A venture capital fund begins life when a group of individuals (GPs) raises capital from a passive group of investors referred to as LPs. Over time, the capital is then invested in portfolio firms identified by the GPs. LPs are legally barred from taking any active role in the security selection or the day-to-day management of portfolio firms.

VC contracts generally take the form of a pre-specified capital commitment. That is, rather than writing a check upfront, LPs are obliged to supply capital over time as requested by GPs. This process of “capital calls” occurs until the entire commitment is exhausted. Funds are set up with a finite life, typically ten years. As the investments evolve and mature, payoffs are divided between the GPs and LPs according to the terms set out in the fund charter. One standard contract is called the “two-and-twenty” in which GPs earn a 2% management fee as well 20% of profits (referred to a “carry”). VC contracts may also specify a clawback provision, which ensures that GPs only retain this carry when the LP’s overall return exceeds a pre-specified (usually low) hurdle.

The Preqin data considered later reports the net return to LPs. More specifically, it reports the internal rate of return of all cash flows of LPs to the VC fund.

4.1 Hypothesis Development

The key prediction is that given a positive shock, low-quality agents enter on both the supply side and the demand side. In contrast, high-quality agents are always active. Therefore, dispersion in agent quality is procyclical. This heterogeneity is likely to manifest itself in returns as portfolio firms mature and uncertainty is resolved.

Hypothesis 1 (Portfolio Firm Returns). Increased entrepreneurial heterogeneity during hot markets increases the cross-sectional dispersion in portfolio firm returns.

The above effect would be muted by diversification if the unit of observation is at the fund level. Venture capital funds are, after all, portfolios. On the other hand, noting that the model makes a symmetric prediction about the quality of supply side agents, we have the following hypothesis.

Hypothesis 2(VC Returns). Increased heterogeneity in VC quality during hot markets increases the cross-sectional dispersion in returns across VC funds.

4.2 Sample Selection and Variable Definition

To test the model, I use VC return data from Preqin. As previously mentioned, these data are at the fund level. More specifically, Preqin reports the IRR of cash flows to LPs, which I use to test Hypothesis 2. Preqin has no information regarding the performance of individual investments within a VC fund’s portfolio, and so this database is unsuitable for testing Hypothesis 1.

Preqin also reports a vintage year, which is defined as the first year in which any portfolio firm investment is made by a particular fund. The sample begins with the vintage year 1980. Given the life cycle of a VC fund, I truncate the sample at 2004 to limit reliance on “interim returns” based on self-reported valuations rather than realized returns. This leaves 1,071 VC funds between 1980 and 2004 with return information. I collect information on fund’s name, its vintage year, and the identity of the parent firm issuing the fund. The 1,071 VC funds are associated with 504 unique parent VC firms. Finally, I collect information on the fund’s abnormal return, which is Preqin defines as the fund’s internal rate of return minus the rate of return on a contemporary benchmark portfolio of VC funds with the same industry focus and specialization. Abnormal returns are available after 1982, and are available for 905 of the 1,071 observations. Although the years 1980 and 1981 lack data on abnormal returns, I retain them in the sample for use in two classification schemes described below.

The key step is to classify whether a fund is offered by a new entrant or not. As a first pass at this identification, I calculate founding date by searching the database for the earliest observation of any fund offered by the same parent VC firm. The fund is considered a new entrant when the parent firm’s founding date equals the fund’s vintage year.

This classification scheme is subject to a potential bias near the beginning of my sample. Because the database is truncated at 1980, some VC firms in the early 1980s may be incorrectly identified as new entrants. To investigate this further, for all VC funds raised in the first seven years of my sample, I performed web and Lexis/Nexis searches to determine if the firm was associated with a previous VC fund. This exercise was typically straightforward, as venture capital firms tend to leave a record of their historical evidence. The majority of cases were not reclassified. However, there were nine clear-cut cases in which I was able to find evidence of a previous VC fund by the same firm. Other cases involved judgment calls. For example, a firm might have had a closely related predecessor VC firm which closed venture capital deals. In a few such cases, I ultimately changed the assignment from new entrant to incumbent based on the weight of the evidence and the spirit of what is meant by new entrant. In total, fifteen firms near the beginning of the sample have been reclassified. Appendix A lists these observations. These reclassifications have very little effect on the empirical results.

As a first look at the data, Table I lists the venture capital funds by year, along with the first four sample moments of abnormal returns.

Note the extreme amount of skewness and kurtosis evident in Table I. This presents a challenge because outliers in the right tail have a potentially large effect on the sample moments. I take five steps to mitigate the impact of these outliers. First, as previously mentioned, I employ abnormal rather than raw returns. This tempers variability related to overall venture fluctuations. Second, in what follows, I winsorize any observation with a return exceeding 300%.¹⁰ Third, as an alternative measure, I use the log of 100% plus abnormal returns. By moderating extremes, logging variables has a substantial effect on standard errors. In what follows, I present all results with both raw and logged abnormal returns. Fourth, I examine the entire distribution of returns rather than focusing on a particular sample moment. Fifth, in the final test, I employ a quartile-based performance ranking released by Preqin. Although quartiles serve as a very crude measure, their insensitivity to outliers confers a distinct advantage.

The mean abnormal return in the full sample is slightly positive, which provides a sense in which the Preqin definition of abnormal return is not completely demeaned. This may reflect censorship of poorly performing funds in the Preqin database.¹¹ Alternatively, it may reflect a difference between in the weighting scheme of Preqin’s benchmark compared with Table I, which is equally weighted. It is shown later (Table V) that the median abnormal return is close to zero in the full sample as well as most subsamples.

The next task is to construct a measure of the heat of the venture capital market. I employ two such measures. The first is based on the number of active VCs, as reported in Table I. (These overall patterns are broadly consistent with Table V of Kaplan and Schoar (2005), which is based on Thompson data rather than Preqin.) Vintage years are defined as hot when the number of active VCs exceeds the average number of active VCs in the previous four years. This classification yields two long hot streaks: 1984–1989 and 1992–2000, with the rest of the sample cold. This metric is undefined for 1980–1983 (the sample’s first four years) which I manually define as cold as they have the lowest activity of any years in the sample.

The second definition of heat is based on dollar values rather than number of active VCs. For each vintage year, I calculate the total dollar value of all funds raised. I then proceed as above, comparing this value to the average of the previous four years. These two definitions agree on all years except 1992 and 2001. Preqin defines vintage year as the first year in which the fund made an investment in any portfolio firm. So, for example, VC firms of vintage year 2001 likely had legal incorporation and fundraising in 1999 or 2000 before the dot-com crash. The Preqin database lacks precise date of incorporation which would permit a more nuanced definition of heat.

4.3 Main Empirical Results

For the first set of tests, I divide the sample into various subsets, categorized by i) whether the parent firm is a new entrant or not, ii) the heat of the vintage year, and iii) the heat of the year in which the parent was founded. The model predicts more dispersion in the returns of new entrants in hot markets.

Subsample moments are reported in Table II. The mean differences are small and are typically insignificant.¹² This result is unsurprising: because VC returns are very noisy, subsamples of modest size are unlikely to detect small differences in mean. Turning to the second moment, however, the evidence is more clear-cut. VC funds are much riskier in hot vintage years than in cold years (19.6% versus 36.7%; p-value  < 0.01).¹³

Table II, panel C and D, also suggests that VCs founded in hot periods have more dispersion than those founded in cold periods, which seems consistent with Hypothesis 2 at first glance. However, classifying according to the scheme above has an important limitation. The hot-founded subsample contains a disproportionate number of funds which also happen to be operating in hot markets. (It not uncommon for a VC firm to originate in a hot market and then die.) Yet, as the evidence already indicated, hot periods are associated with more risk. Thus, we are unable to conclude that this difference in Table II, panel C solely reflects the hot-founding categorization.

To address this limitation, the next test divides the sample more finely. I consider a 2× 2 breakdown according to the heat of the fund’s current vintage and of its parent firm’s founding. For example, a fund may currently be operating in a cold market, whereas its parent was founded in a hot market. The returns in these subsamples are summarized in Table III. As before, most of the mean comparisons are statistically insignificant. These (unreported) t-tests are of limited power given the small number of VCs founded in cold periods.

Turning to the comparison of cross-sectional standard deviation, during hot markets there is more variability among firms founded during hot markets than cold markets. This difference is large, both economically and statistically (11.8% versus 38.7%, p-value  < 0.01.), consistent with the model’s predictions. On the other hand, the comparison is complicated by the fact that the “Hot-Vintage-Hot-Founding” group is heterogeneous. It contains some funds whose parents were founded in another, earlier hot vintage but also includes other funds which are new to venture capital. In contrast, the “Hot-Vintage-Cold-Founding” firms are all seasoned by definition. Their survival to another period is noteworthy. This large difference in Table III may partially reflect the safety that comes with seasoning.

To investigate this issue further, the next test involves dividing the sample first according to whether the VC firm is a new entrant or not, and then according to whether the fund under consideration is in a hot vintage or not.

These comparisons constitute the main results and are summarized in Table IV. The first observation is that there is generally heightened variability in hot markets. This regularity was already noted in previous tables and is a very strong feature of the data, as seen in the extremely low sig-F values throughout Table IV. Consistent with the model, the volatility gap between cold-and-hot markets is larger for new entrants (39.2% versus 15.8%; p  < 0.01) than for incumbents (34.5% versus 21.9%; p  < 0.01). Thus new entrants in hot markets are 2.5 times as risky as their cold market counterparts, whereas for incumbents this ratio is 1.57 times. The second definition of heat leads to similar comparisons. Overall, new entrants in hot markets are associated with more dispersion than any other subsample, consistent with the model.

As previously mentioned, one drawback of the tests summarized in Tables II–IV is that sample moments are sensitive to outliers. The next test instead employs a quartile performance ranking released by Preqin. For each observation, Preqin places the fund into one of four buckets according to relative performance based on vintage year and “where possible, funds which adopt similar strategies and a similar geographic focus.” As is the case with their abnormal return measure, the benchmark comparison set and precise methodology is proprietary. While this measure is crude, it is insensitive to outliers. This test has an additional advantage in that the sample size is slightly larger than in previous tests: Preqin assigns performance rankings for some firms for which it does not release an abnormal return measure.

Table V summarizes the distribution of rankings earned by each subsample type. New entrants in hot markets are more likely to earn the worst quartile ranking Q4 (30.1%) than any other subsample. Incumbents in hot markets are unlikely to earn this poor designation (20.6%). In fact, the distribution of incumbent rankings seems relatively insensitive to markets conditions overall, as indicated by the fact that rows (c) and (d) are fairly similar, and rows (g) and (h) are very similar. New entrants in cold markets are the most likely subsample to earn the top performance ranking Q1, as seen in rows (c) and (f). These top marks are consistent with the high average quality $πmin+12$ of new entrants posited by the model during cold periods.

The final test aims to obtain a better sense of the full distribution of returns within each of these subsamples. Table VI reports the decile breakpoints for VC fund returns. The median abnormal return is near zero in all subsamples. As expected, the left and right tails both become more pronounced during hot markets.