On Knowability
Investors agree on the most important factors that determine an investment’s success or failure: what multiple did you pay, at what multiple did you sell, how did profit change in the interim, and how long did you hold.
I differ from the orthodox view not in my evaluation of importance but in my evaluation of knowabliity.
I think that humans are far less good at predicting future profit growth, future exit multiples and even hold periods than most people suspect. Few hypotheses have been more proven in social science than the inferiority of human predictions to simple mechanical prediction algorithms, yet many investors still place an inordinate weight on their own predictive models.
Donald Rumsfeld famously divided our understanding of the world into three categories: known knowns, known unknowns, and unknown unknowns. Stealing from Rumsfeld, I have developed a similar framework for investing. Investors can divide the inputs into an investment process into three categories: the important knowns, the unimportant knowns, and the important unknowns.
Figure 1: Importance & Knowability
To effectively synthesize historical data and predictions about the future, investors need to multiply the importance of each variable by its knowability. The only way for fundamental investors to get at knowability is to systematically track the predictions they make about earnings and multiples and then compare those predictions to reality years in the future. This requires discipline and foresight and a long track record.
I posit that investors that looks at this data will find the following:
Future earnings follow a random walk. Predicting that every company in the stock market will grow 3% per year has historically been a more accurate method of predicting earnings than relying on the consensus of investment analysts.
Investment multiples mean revert. Investors can roughly assume that valuation multiples revert 25% towards the mean every three years. So if you are trying to predict the EV/EBITDA multiple of a company that trades at 18x EBITDA when the long-term market mean is 10x, 16x would be a relatively good prediction three years hence. This predictive method is superior to relying either on current multiples or peer multiples.
Future financials and exit multiple are equivalently important to entry financials and purchase price. They are, however, significantly less knowable. In fact, many people confuse historical knowledge with predictive accuracy, thinking that the more they know about a given business the better they will be able to predict the future of that businesses’ financials and valuations.
Historical financials, purchase price, and current capital structure are very important to the ultimate outcome of the investment and also happen to be completely knowable. But in order to interpret this historical data, we need to evaluate each piece of data by how predictive it is of future prices.
This is where linear regression models and machine learning analysis of historical equity data is helpful. We can separate relevant data (e.g. EV/EBITDA multiples) from irrelevant data (number of employees, years of tenure of the CEO) and weight that data by predictive power (e.g. purchase price is more predictive than historical debt paydown which, in turn, is more predictive than year-over-year change in net income). Quantitative investors and academics produce voluminous research on these topics, and the best investors need to understand and apply this work in order to compete in the arms race for investment knowledge.
Brian Chingono and I have laid out the key insights from both linear regression models and machine learning to optimize the Verdad strategy, and we believe that our robust analysis of historical data allows us to have a superior interpretation of historical financials and price data than our competitors, particularly in the world of leveraged equities. We hope that if you have not yet read out papers in full, you will take the time to click on the links and learn more about how we approach investing. See Evans et al,"An Examination of Economic and Statistical Approaches that Address Sample Selection Bias, Inaccuracy, and Optimism in Analysts’ Earnings Forecasts,” 2012 for more on predicting earnings. Email me for the Verdad research on mean reversion of multiples.