Proactive Forecasting

Joel Greenblatt in both ‘You Can Be a Stock Market Genius’ and ‘The Little Book that Beats the Market’ follows a similar investment generation methodology. He finds baskets of companies which, as a group, tend to outperform the market. He then digs into those baskets with fundamental analysis to juice the returns further, with the knowledge that even if here to add little or no value in the fundamental analysis process, that he still has positive expected returns to back him up because of the risk/reward properties of the baskets being looked at. His “baskets” included spinoffs, partial spinoffs, stock recapitalizations, merger securities, and stocks which are cheap and good, where cheapness is defined by earnings yield and goodness is defined by return on invested capital. For one reason or another, all these groups taken as a whole outperform, so even if he were to pick stocks at random from these lists, under the right set of conditions, he would still outperform.

So the way I see it, his methodology is able to take advantage of the benefits of both quantitative analysis and fundamental analysis.

Quantitative analysis is very good at using large amounts of historical data to back-test things which we may intuitively believe to be true. In this fashion, it can be very helpful as a check, and it can help us form a more reasonable expectation of the sort of returns we can expect from a particular situation over time.

Fundamental analysis is less useful for back-testing because proper analysis requires so much time, but it can reach a depth of understanding which just isn’t possible with quantitative analysis.

Greenblatt (and Pzena), by leveraging both, haven’t done too badly.

The goal of their forecasting is to find pockets of companies which tend to outperform the market. The thing which should be noted, though, is the fact that all their predictor variables are pre-visible—they are all things which are known with complete certainty at the time of investment. For example with the magic formula, the ttm ROIC and earnings yield are by definition already known—there is no uncertainty that those numbers aren’t true.

If the only goal of forecasting is to find groups of companies which tend to outperform, why should we constrain our predictor variables like this though? I would introduce the notion of “cost of error in my predictor variables” as well as “predictability of my predictor variables.” In this context, I would claim the following:

{Usefulness of an input variable} = f({ability to know input variable}, {ability of input variable to predict output variable}, {cost of error if input variable’s actual value deviates from expected value}).

What typical regressions assume, in this paradigm, is that the cost of not knowing what our input variable’s actual value is is infinite. This forces us to make forecasts entirely on the basis of past data. I could see some value though in including input variables which are forward looking—I might not know what their value will be exactly, but if I know that I can predict those input variables with a good level of confidence (through a lot of due diligence, for example), then those input variables could be a lot more useful than input variables which strictly look to the past. While I'm on the subject of typical regressions, I'd also like to add that most people tend to get more than a little bit lazy in their data collection. Why should I constrain myself to variables that I can easily get, or that I can easily quantify? This misses out on the whole notion of cost. There are a lot of "fuzzy" variables that could provide wonderful insights to any quant model, if only someone would just go and do a little more digging-- be a little more subjective-- and stop being so damn traditional for once.

Anyways I digress.

If I find an input variable which I think can predict with a good level of confidence future returns, but I’m not 100% sure what the input variable’s value will be (for example, next year’s earnings), then {ability to know input variable} decreases but {ability of input variable to predict returns} increases. As long as the cost of deviation is low, I could very well favor this input relative to historical inputs.

This takes Greenblatt’s methodology one step further and completes it. In this context I would run screens like the following: find all companies experiencing massive EBIT growth relative to their current EV/EBIT, and which subsequently are able to maintain EBIT growth over the next four quarters which is at least twice the level of the EV/EBIT. See how these companies have performed over the past 20 years. Analyze the distribution for patterns—are there periods of time where this sort of methodology fell out of favor? Do the losers exhibit a certain quality in a non-random way? If these companies dramatically outperform the overall market, then I know that if I were to screen for companies with massive EBIT growth relative to EV/EBIT, and I was able to predict with a high degree of confidence that that EBIT growth would hold up for at least a year for some subset of this group, I would probably consider constructing a trading strategy around this.

This again uses quant in addition to fundamental analysis, but brings them together much more tightly. This could be useful for the idea generation process. It requires a high level of discipline in the stock picking process, in a similar fashion to Greenblatt and Pzena. It might make a few people some bucks.
-Dan