Is Venture Capital Intrinsically Cyclical?

In technical language, public-market returns generally do not have meaningful autocorrelation.

Private markets are different. Private equity, real estate, venture capital, and illiquid hedge fund strategies often show serial correlation. One month’s return predicts the next month’s return more than it should. This is not a new observation. In 2004, Getmansky, Lo, and Makarov showed that hedge fund and alternative investment returns are often highly serially correlated, and they connect that serial correlation to illiquidity exposure and smoothed reported returns.

The real estate literature got there even earlier. Geltner’s work on appraisal-based returns from 1991 is the canonical reference that suggests appraisal series can be smoothed, which significantly understates volatility. [2]

The modern private-equity critique is the same basic idea in allocator language. Private-market returns can look artificially polite because assets are not marked to market continuously. Cliff Asness, the founder of AQR—one of the flagship systematic trading firms—has written that private equity’s artificially smooth returns make the asset class difficult to model, and that investor demand may partly reflect a preference for smoothed return streams. [3]

Enter: “unsmoothing”. The standard unsmoothing move is to estimate how much autocorrelation exists in a private-market return series, then construct a synthetic return stream with the same general economics but less serial dependence and more realistic volatility. In the usual framing, unsmoothing is an epistemic correction. The reported series is too smooth, so we infer the more violent series hiding underneath it. There is a real market price somewhere but reported NAVs are just slow to catch up, because of stale appraisals or managerial malfeasance.

The implication is that in a perfect world of fast, direct marks, private markets should not need to be “unsmoothed”. Venture is stranger than that.

AngelList gives us a cleaner test

AngelList has price-per-share data on thousands of startup investments, administered through a consistent valuation policy. For this analysis, we wanted a broad, seasoned, representative portfolio with enough history to observe cycles. So we built a simple portfolio by putting $1 into every Seed investment on AngelList from 2015 through 2017, and then measured gross returns each month from January 2018 onward.

That gives us almost 1,500 investments. This is not meant to represent a directly investable product; we use gross returns because we want to look through investments made by funds and understand underlying asset-class-level performance. This portfolio is not a complete picture of venture capital; AngelList was and continues to be underrepresented in biotech startups, and it misses companies like Anthropic that now represent a huge fraction of the residual value of the AngelList platform. Although our demonstration portfolio is large, it’s possible that a larger portfolio with more diverse exposure could have different dynamics.

The portfolio compounds gradually, then vigorously through the 2021 to 2022 boom, then mostly plateaus around 5x the initial 2018 mark.

The monthly return distribution is interesting, because the left tail appears to follow central limit theorem aggregation, while power law of winning early-stage investments retains a more explosive right tail:

The observed monthly returns are positive on average, with a high share of up months and low measured volatility. On the surface, this looks like an unusually well-behaved compounding engine; if we were getting a random draw from this distribution each month we would be very happy.

We are not getting a random draw from this distribution though, because of autocorrelation.

Return persistence on AngelList Data

We fit an autoregressive model to the timeseries data. The best-fitting model for predicting monthly log-NAV changes was a two-lag model, with each month’s portfolio change seeing large, nearly equal contributions from the past two months:

The partial autocorrelation plot shows the first two lags as significant after Bonferroni correction; two lags also minimize AIC and BIC.

The number of lags is not the same as the number of months a trend persists. In an autoregressive process, prior months keep influencing future months through later months. A good month affects the next month and the month after that; those months then influence later months, and so on.

Under this model, the implied new-information share each month is roughly:

So only about 23% of the monthly move behaves like fresh innovation. The rest is persistence from prior months. And again, this is the important part: the usual explanations for persistence are much weaker here. AngelList data is monthly, the marks are consistently administered, and the dataset is built from investment-level back-office records rather than manager-selected fund performance reports. But the autocorrelation is still there!

That pushes the interpretation somewhere more interesting: what if venture marks were discovered in sustained waves, as opposed to merely being slow to update?

What unsmoothing shows here

Applying our two-lag model to the returns data produces a much more volatile synthetic return stream.

The red line is not literally what this portfolio would have traded at every month. Nobody was making a two-sided market in a thousand Seed investments. The red line is a diagnostic. It asks how much volatility you need to introduce before the return stream stops behaving like a smoothed private-market series and looks more like a volatile public market return stream (the X and Y axis are identical to the last histogram to facilitate comparison):

The summary is blunt:

Observed annualized volatility is 9.3%, but after unsmoothing, it is 30.7% [4]. Up months fall from 71% to 58%, much closer to what we would expect in public markets. The portfolio goes from looking like a polite compounding machine to something much closer to what intuition says venture should be: high-return, high-volatility, regime-dependent risk.

In the usual private-market setting, unsmoothing is often treated as a correction to reported NAVs. Here, I would interpret it somewhat differently. I believe that the unsmoothed series is a diagnostic for the sheer amount of cyclicality embedded in the venture price-discovery process.

I want to push back against facile explanations here. The unsmoothed series is not the “real” monthly venture price, and the observed series is not some dumb accounting fiction either. The interesting thing is the gap between them. It tells us how much market-regime persistence is being compressed into a smoother observed path.

Public, liquid venture could be weird

The unsmoothed values may be representative of what venture capital would trade at it in a liquid market, since they “reverse engineer” a martingale, making them closer to the kind of price motion we would observe if venture portfolios were publicly traded. That means if venture assets were traded continuously, we could expect them to fluctuate around their NAVs with high volatility.

That matters for new publicly traded venture vehicles like RVI [5]. If a vehicle owns venture assets but trades daily, the market price may swing far above and below reported NAV just as the unsmoothed series did in our analysis. That can look irrational if you treat NAV as truth, but it looks less irrational if NAV is a smoothed estimate of a much more volatile economic value.

A publicly traded venture portfolio should not be expected to behave like a private NAV with a ticker. The ticker imports daily market psychology into an asset class whose underlying value is revealed in bursts.

The two-lag model is highly suggestive

The best autoregressive model uses two lags with large similar weights. I think this specific result has additional significance.

Registering a markup, markdown, or wind-down in month X versus month X+1 is partly arbitrary. A round closes on the 29th instead of the 3rd; a shutdown gets processed this month rather than next month. That kind of boundary noise will create spurious short-term mean reversion. One month catches the big markup and looks good, but the next month looks worse because it did not.

A two-lag model with large equal coefficients will effectively try to eliminate this short-term mean reversion by “netting out” boundary effects. But if we’re already netting out spurious trends, the implication is that monthly reporting is the limit of useful data accuracy in venture capital. Weekly or daily marks would only create more precise timestamps on inherently lumpy information, which is not the same thing as greater truth.

This matters because one traditional argument for unsmoothing is that private-market data updates too slowly. But if the monthly data we’re using here is sufficiently high fidelity to show boundary effects then the lesson is different. Venture’s return process looks to be autocorrelated even when the data is timely, because startup value is discovered through sparse, correlated events.

Why this matters for vintage years

Intrinsic persistence gives a quantitative explanation for something venture investors already believe: vintage year matters a lot. If private-market smoothing were only an epistemic problem, then over long enough windows it should mostly wash out. Marks may be stale for a quarter or two, but eventually the truth will out.

However, if venture returns have deep cyclicality, vintage years differ because different cohorts are born into different regimes. A 2015 Seed company and a 2021 Seed company are not merely separated by six years. They were financed under different capital costs, different follow-on environments, different exit expectations, different public comps, different talent markets, and different beliefs about what growth was worth.

One of the takeaways is that every time we look at venture data we should always be benchmarking what we’re looking at against the rest of the market at the time. A 3x multiple on Seed investment after three years looks good but not great for a 2015 investment, but terrific for a 2021 investment.

The converse of this takeaway also holds: any venture benchmarking or comparisons across time may be inherently flawed. For instance, comparing a manager’s Fund I vs Fund II in isolation may not provide any real insight, because those funds would have experienced different regimes. Furthermore, data that can only be observed over very long windows, like distributed capital, may be useful descriptively and but have limited predictive power. By the time the result is visible, the regime that produced it is gone.

The stronger conclusion: venture value arrives in waves

Dick Kramlich said venture capital “was and always will be” super cyclical. By fixing the two big problems of private-market reporting - timeliness and manager discretion - I think the AngelList data gives us a new reason to take the “always will be” seriously.

Capital availability, follow-on rounds, exit windows, public-market comps and investor risk appetite are all correlated. When the market decides that growth is worth more, many startups become easier to finance over a period of months. When the market decides that growth is worth less, many startups become harder to finance over a period of months. And when a world-sweeping change like generative AI hits, every startup is affected.

Those are not independent company-level events. They are regime-level events expressed through company-level marks.

This is why I think the “delayed information arrival” explanation for private market autocorrelation misses the point. In venture, the lumpy events are the information. A financing round is not a stale memo about a price that existed three months ago. It is price discovery. A shutdown is not a delayed observation of a continuous market price. It is the end of the price process. A hot follow-on market is not merely a reporting artifact: it changes which companies survive, how much runway they get, what prices investors accept, and what exit outcomes become plausible.

We believe that venture capital does not just “have” lumpy observations. We believe that venture capital is lumpy, correlated, self-reinforcing observations. Our results suggest that there’s not a better, more accurate and timely dataset than what we have access to at AngelList. The implication is that cyclicality is not sitting on top of venture capital: it is inside it.

For investors, this cuts two ways. If you could predictably time venture cycles, persistent regimes would create opportunity. More realistically, most investors cannot do that, which makes diversification across time more important. Broad exposure across companies helps. Broad exposure across vintages may matter just as much.