Early Stage vs Late Stage: Assessment, Prediction & Fluidity In The Equity Chain
Making an investment decision involves both analysis and intuition. These decisions contain an irreducible intuitive component, but when early stage investors rely largely on intuition, growth stage and later stage investors will increasingly add rational, objective elements and models to make their decision. It’s very natural indeed: the level of information differs largely between the stages. Late stage investors will have much more information at hand about the company and its market.
A distinction should be made between two types of information that are available to investors. Daniel Kahneman and Amos Tversky in their article Intuitive Prediction: Biases and corrective Procedures (2013) describes the differences between singular and distributional information.
Singular information, or case data, consists of evidence about the particular case under consideration. Distributional information, or base-rate data, consists of knowledge about the distribution of outcomes in similar situations.
Early stage investors will tell you that they have almost no singular information. If the go-to-market is not yet done, there is no data to analyze. So the internal view, using singular information, will be focused on the founding team. Investors have to make a leap of faith. And there is very little to say in that regard — without data, no formal analysis is possible.
But what about distributional data? Couldn’t early stage investors use the outside view and look for comparable cases and base rates? You could argue that, with the exception of certain companies trying to define brand new categories, most companies are targeting existing markets where at least some data could be found.
And yet, it’s not uncommon to hear professional early stage investors, specifically at the seed stage, asserting that given the level of uncertainty, it would be a waste of time to make investment memos or to organize calls with professionals working in the market since they would be biased against change.
Opponents of these methods (or the lack of) would argue that the different risks could be identified and analyzed.
Early stage investors would (I say would, but this is a quite common discussion) typically answer that “if you focus on the risks you would never invest in young tech startups: you can always find reasons things can go wrong — people give up, products may be disappointing, competition could execute faster, consumer behaviors could be mis-anticipated, technologies (or science for deep tech startups) may not work, team members may get misaligned, founders could fail to raise enough money to survive, markets could be smaller than expected, unit economics could go sideways due to increased cost of acquisition or competitive pressure on price, etc.“ And that would be a fair point.
On top of that how could investors approach the risks in such situations? It’s almost impossible to assess properly every single one of them, the margin of errors would be too big. Attempting to add slippage factors wouldn’t be much more adequate, since it would keep a value close to the initial prediction or make it very subjective — this is by the way a good example of the difference between accurate and precise, investors would get a precise number on a preconceived opinion, thus have something precise but most probably not accurate. The solution (as you have guessed) would be to take the outside view, exactly like you would do to estimate someone’s life expectancy. You would start by looking at the average life expectancy (and more specifically for a given gender), instead of just trying to guess widely focusing on the apparent vitality & resilience of the subject. And to assess the risk of a random seed stage company picked by a professional seed investor, you could use a proxy such as the conversion rate from seed to series A and use a database such as Dealroom (and make a deep, long, honest self-reflexion on the bracket you put yourself into).
Thus you could say that depending on the skill of the professional and her level of conviction, the risk of failure before the next round is between 60% and 93%. This wouldn’t help you pick a specific company, but this would be helpful to build a portfolio model for instance.
But in fact, there is a continuum between the internal and outside views. Investors could increase the granularity and look for base rates such as the risk of failure for specific subsectors (eg: they could favour enterprise companies or fintech companies, if they find more promising data, all things being equal).
This mechanism of comparing opportunities (which is the core of finance) and increasing granularity would progressively build a bridge between the outside and the inside view and perfectly describe what early stage investors actually do: they compare a case with all the others met before to converge progressively to a solution (which isn’t without making us think of the groping mechanism described by the Walrasian auction). It’s also the reason why first investments are tough, the basis of information being too small for this exercise. When investors meet founders and assess opportunities, they are thinking something like “I would say that she is among the 5% of founders that impressed me the most“. Said another way investors predict by “matching prediction to impression“ (which raises another topic of the premium given to good storytellers) and groping from meetings to meetings.
So naturally, almost intuitively, all investors use some kind of distributional data. Yet, by not pushing the outside views further to assess risks and improve their prediction ability, early-stage investors are putting a sword of Damocles over their head.
In fact, the methodologies of risk assessment and prediction are intentionally ignored by most early-stage professionals because they think it’s structurally impossible: singular information is incomplete, distributional data are largely nonexistent or difficult to get and the environment is very shaky. These are the rules; intuition, faith and sweat are what matter. And the game is played for a clear reason: the magnitude of gain in the case of success is compensating the risks. Nothing new under the sun, the expected payoff for a given company is nothing else than the probability of success multiplied by the potential payoff (and in the portfolio, it’s the sum of these multiplications). The probabilities are hard to quantify and they are certainly low, but the payoff is very high.
The thing is: as startups go up in the financial chain, this magnitude of potential gain tends to decrease since the range of exit values stay relatively constant, while the price paid increases round after round. So at some point, later-stage investors need to be much more diligent about the risks and get more precise about forecasts, explaining the need to properly consider the outside views. Happily, the inside view too is improved by many sources of data that were not available before: growth investors and later-stage investors get real business plans, the evolution of a top-line and the associated margin, costs of acquisition (or at least sales and marketing costs), they also get usage data, such as level of engagement, churn and user behaviours cohorts by cohorts. But they also need to be much more attentive to the market for exits, looking at M&A activities and multiples for comparable companies in the public market.
Using these concepts of inside and outside views and having in mind the dynamics across the equity chain (the tension between potential payoff (the magnitude) and the risks of failures) could clear up many misunderstandings and help improve the fluidity within the equity chain by predicting potential frictions along the way:
- all tech investors are biased towards looking for very high upside potential given the high level of failures, the only way rational way of doing tech deals with lower upside potential being a strong conviction that the risks involved are way lower than peer companies,
- the market for exit needs to be large and dynamic enough to allow late stage investors to take the risks of funding high priced scale-ups (with significant goodwill), which is one of the reason why the chain of financing get stuck during crisis, when exits are frozen,
- early stage investors have the luxury of not asking for precise models such as a proper business plan or market sizing because there activities are grounded on uncertainty, high failures rates and colossal payoff potentials (it’s both a luxury and a curse, depending on how you see things),
- increasing valuation overtime leads late stage investors to look not only for big enough upside potential but also for more clarity about these payoff potentials and the attached risks involved. Thus, they will ask for more data about a company (the inside view), spend more time analyzing the market (economic -demand and competition- and financial — comparable equity stories and exit events — the outside view) and build more models (to balance properly the inside and the outside views),
- these dynamics make it essential for founders to change drastically their fundraising materials and overall approach — you can’t ask late stage investors to make blind leap of faiths, it just does not work at their stage,
- founders of startups that do not match certain patterns get frustrated because they cannot easily finance their ventures but the explanation lies largely on the lack of distributional information (and even early stage investors which does not formally use the outside view will get concerns about the magnitude of the potential payoff, which is a way of saying that they are not sure that they can keep their generic range of exit potential used, since the reference class seems too different for a legitimate use of groping),
- during a crisis, investors need to get new distributional data to make good decisions: companies do not uniformly react so you need both to reassess all your references classes and how all classes are impacted,
- to get a fully functioning equity chain, investors need to understand each other constraints and appropriate practices, and collaborate on the complementarities: early stage investors can share their perspectives on how the team works and how well they resolve issues (inside view), on what struck them compared to other teams in that stage, and their intuition on changes about technologies, consumer behaviors and products (let’s call that proto outside view), while trying to anticipate later stage frameworks and requirements (to select beforehand the fittest startups to match their investment criteria and to help the startups get there and prepare the adequate materials); later stage investors can share their perspectives on the outside view (both at the company level and about the macro-environment, both economic and financial), while trying to benefit from insights from earlier stage investors to fine-tune their hypothesis and best calibrate the residual need for intuition, since every tech companies, however mature, are facing certain degree of incertitude. This is by the way, one of the strengths of our positioning at Gaia Capital Partners, getting the perspectives from all early stage, growth and public market investors.
- every decision relies on intuitive judgements, and they are often biased. The good news is that “intuitive judgements are often biased in a predictable manner“ (ibid). Hence, the problem is not wether to accept intuitive predictions at face value or reject them, but rather how they can be de-biaised and improved.“ And as you can guess it’s all about balancing properly singular and distributional information; the typical mistake extensively studied by Kahneman being the neglect of the outside view.
For the most curious readers, you can find in the appendix below the method recommended by Kahneman and Tversky as an addendum of this article. It’s a five-step process involving: the selection of a reference class, the assessment of the distribution in that reference class, making an intuitive assessment, assessing the predictability (through linear regression) and eventually correcting the intuitive estimate.
A more general discussion could be made about the method, asking why the predictions should be regressive. The procedures recommended by the author will “usually yield conservative predictions that are not far from the average of the class and are very unlikely to predict an exceptional outcome that lies beyond all previously observed values“. The answer to this objection from the authors is that “a fallible predictor can retain a chance to correctly predict a few exceptional outcomes only at the cost of erroneously identifying many other cases as exceptional“.
As I was reading this, I could almost hear the typical discussions between optimistic early-stage investors and more conservative late-stage investors that tend to rely more on distributional data. And unlike the generic prescription of Kahneman saying that taking the outside view does matter, I may suspect that early-stage investors could be often right despite ignoring the outside view, given the instability of our environment (and the fact that distributions follow a very high skewness, so taking a too broad outside view leading to non-meaningful averages could be misleading, as we’ve shown last week).
Addendum - How should we make good decision, a five step process
- Select a reference class: Sometimes it’s straightforward (eg: life expectancy) but there are difficulties: 1.1 it’s always difficult to determine which is the right granularity to be applied and which variables should be selected (eg: gender ? gender+location? gender+location+level of income? gender+location+level of income+ level of study? gender+location+level of income+ level of study + eating habits?, gender+location+level of income+ level of study + eating habits + practice of physical activities, etc.). It should be noted that there is often a trade-off between conflicting criteria, for instance the most inclusive class may allow for the best estimate of the distribution of outcomes, but it may be too heterogeneous to permit a meaningful comparison to the book at hand (there could be a whole another discussion about the non convergence of the parameters leading to the impossibility to establish a generic method and a perfect estimate) 1.2. what can you do when it’s an innovation or at least when the various instances appear to be so different from each other that they cannot be compared meaningfully
- Assessing the distribution for the reference class: Sometimes it’s an easy task — the statistics are available (and it’s tempting — and sometimes relevant — to modify the reference class to benefit from existing data). But more often you have to estimate the distribution on the basis of various sources of information.
- Making an intuitive estimation: It’s the activity that rely on leveraging the singular information about the particular case at hand, which distinguishes it from the other member of the reference class. This is the non-regressive step, done intuitively. It’s what many seed investors do exclusively. The next two steps are designed to improve this estimate.
- Assessing the predictability: It’s the step where you try to see if the type of information available in this case permits accurate prediction of outcomes. Focusing on situations where linear regressions are not possible (ie: when records of past prediction and outcomes does not exist), there are two main possibilities: 4.1. comparing the predictability of the variable concerned to predictability of other variables, eg: it’s less predictable than temperature but more predictable than stock prices. 4.2. asking questions such as: “if you were to consider two startups that you are about to fund, how often would you be right in predicting which of the two will generate a better return of investment?“ It should be noted that estimates of predictability are not easy to make — people are subject to hindsight fallacy, which leads to an overestimate of the predictability of outcomes, and to an availability bias, which leads them overly weight recent or memorable information.
- Correcting of the intuitive estimate: “To correct for nonregressiveness, the intuitive estimate should be adjusted toward the average of the reference class“. For the more curious readers, it would mean that the distance between the intuitive estimate and the average of the class should be reduced by a correlation coefficient. Taking an example, the concrete method would be the following: the investor prediction of the exit value is 200m€ and, on average, the exit value of companies who benefit from venture capital money is 100m€. Suppose that the investor believes that he would correctly orders pairs of startups by their future exit value on 75% of comparison. In this case, the predictive accuracy (a) can be estimated using the correlation coefficient (p), using the formula a = 2p — 1, in this case, a = 2*75% — 1 . = 0.5, and the regressed estimate of exit value would be 100 + 0.5(200–100) = 150m€