Machine Learning and Discretionary Investing
In my last article, I discussed four ideas that I find especially useful for understanding machine learning (ML) and how it compares to direct human reasoning. In this article, I want to discuss some strategies, mainly related to discretionary investing, that I think properly leverage the advantages of human reasoning.
The Uses and Dangers of General Intelligence
As I described in the article, the main area where I see humans outperform (for the foreseeable future) is in “general intelligence”: having a flexible mind that can look at drastically new situations and make sense of them. ML systems, in contrast, are trained in specific, controlled settings and can often perform very poorly when even small aspects of the inputs are changed.
In principle, one way that humans can improve the performance of purely software-based decision making is by noticing cases where the examples differ in important ways that are hidden from the software.
As one example, a stock screener that I look at frequently has recently recommended stocks that I know have run up in price due to a buyout offer. The stocks are now trading at almost exactly the buyout price. I also know that the screener is using purely technical signals, so it is unaware of the buyout information. It is predicting a move higher in the stock, mainly due momentum effects, but with general human intelligence, I can spot additional information that tells me these predictions are overly optimistic.
In practice, however, we have to be extremely cautious about overriding the software. Even in the case above, it is possible that the price will move higher due to an even larger buyout offer. And it may be the case that, even if the software doesn’t know which of the examples that it learned from were due to buyout offers, since those examples were included in the training data, they are still informing its decision making. (The proper thing to do— rather than simply overriding the system — would presumably be to add buyout information to the training data, retrain the system, and then see what it thinks of these stocks.)
It is important to keep in mind the example that Joel Greenblatt gave us when he allowed his clients to override his “Magic Formula” screen by selecting the stocks they preferred from those that it selected. By incorporating their own knowledge and intuition, nearly all his clients were able to turn a portfolio that soundly beat the market into one that underperformed. They might have thought that they were improving the results, for example, by not buying a stock that was the topic of a recent spout of bad news. However, the software might have been trained to take advantage of exactly this situation: if many past examples had their own bad news that caused the stock price to drop too far (an overreaction), then the software would have rightfully learned that these were buying opportunities, even without knowing the specific reasons why each particular stock was cheap at the time.
In short, applying your general intelligence is not as just recognizing that “the software doesn’t know about” something so therefore we should override it. Most things that it doesn’t know about have no significant impact on what the predictions should be, and even for those that do, it is hard for us humans to know how much or even which way we should adjust them in light of all the other information already known to the software about the example.
Below, we will look at some strategies that incorporate a human’s general intelligence in a way that I think is sensible.
Seeking Out Unique Opportunities
The easiest situation in which we can feel comfortable overriding the software is with examples that are so unique that we know there are not enough prior examples for the software to have learned how to handle them. Of course, every real world example is unique in some way, but here, I mean those that are unique in ways that obviously make a large difference to the prediction.
The investor that I most associate with this idea is Peter Lynch. Most people associate Lynch with the idea that you should “invest in what you know”. However, in reading more details of his investment career, the quote that I most associate with Lynch is “the person who turns over the most rocks wins.” That is aspect that I am referring to here.
Lynch turned over a huge number of rocks and found many unique companies to invest in. My favorite example was his investment in “thrift conversions”, which were community banks becoming public companies. Due to the incentives of the bank officers, the shares in the new company were almost always significantly underpriced. In order to get in at the IPO price, however, you needed to have an account at the bank, so Lynch had his assistants drive around the country opening accounts at all the community banks that were undergoing this process.
Lynch also had large investments in Fannie and Freddie. They were unique due to their government backing, which allowed them to borrow money at a lower rate than any competitor. As a result, they could make money with a lower interest rate spread than anyone else. Lynch called Fannie Mae “the best business, literally, in America.” They were truly unique opportunities.
Another favorite of mine were regional restaurant chains that were expanding nationally. We know that, on average, high growth rates quickly revert to the mean (see the charts here, for example). However, if a company already has agreements in place to borrow the necessary funds, franchisees already lined up to open the new stores, etc., then the odds that they will hit their planned growth rate is much higher. Hence, if the company is priced at a multiple similar to that of the overall market, it is actually underpriced.
Lynch’s books have many more examples of unique investment opportunities that he discovered by turning over all those rocks. In each case, he had to apply his general intelligence to see why it was a unique situation and why it was significantly undervalued. Those types of opportunities still arise today and we would not expect ML systems to spot them.
Reducing to a Few Key Variables
In contrast to general intelligence, where humans have a distinct advantage, ML has a clear advantage with high-dimensional data, where each example has a large number of attributes that could affect predictions. ML can incorporate all of these attributes systematically to improve the accuracy of its predictions, whereas for humans, extra information usually does not improve accuracy (instead, it primarily makes us more overconfident).
Rather than trying to grapple with all of that information (incorporating it all into, say, a discounted cash flow analysis), an alternative is to seek out cases where it is not necessary to do so.
In investing, this can happen in cases where it is possible to reduce the analysis of the business to only a “few key variables”. While the rest of the available information likely does impact the analysis, it may only move predictions within a small range, and if even the low end of that range is far below the current stock price, then it is not necessary to study further. (Value investors would say that there is a large “margin of safety” in these cases.)
Bill Miller has promoted the idea of finding stocks where only 3–4 variables really determine the value. He is the investor I most associate with this approach, as he is willing to follow the approach into stocks like Amazon that most value investors would never even consider, while still buying classic value stocks (e.g., Teva Pharmaceuticals, today, with forward P/E of 3.4).
Pedro Domingos famously said that “machine learning would not be necessary if humans could see in high dimensions”. If we can reduce the problem to only 3–4 variables, however, it is possible for us to see and understand even very complex relationships between the variables that might be hard for an ML model to fully capture with limited available data.
Superforecasting
As a final example, I want to look outside of investing, at more general examples of making predictions — a.k.a., forecasting.
In his book “Superforecasters”, Philip Tetlock described the general procedure that the worlds best forecasters use to make predictions about world events:
- Identify a class of similar historical examples.
- Calculate the historical odds of the event occurring (the “base rate”), possibly by building a simple statistical model using, say, linear regression.
- Incorporate additional information by adjusting the predicted odds up or down, but only by small amounts (and reluctantly).
For me, it is easy to see this as humans trying to simulate how ML would approach the problem. Step 2 is literally the application of (very simple) ML techniques. Step 3, where information from additional attributes is incorporated, is reminiscent of “boosting” techniques in ML, which continue to improve predictions by incorporating extra variables in just this manner.
However, this procedure also incorporates the unique abilities of humans. In Step 1, we use our general intelligence to identify the set of similar historical examples. In Step 3, we use our general intelligence to identify additional information that is not incorporated into the simple model but likely does affect the true odds. These two parts are not things that ML can do for us. They requires us to draw on our general knowledge of the world and sometimes even our creativity.
Viewed this way, superforecasting incorporates both human and machine intelligence, using both where they are best. Human reasoning is used to identify a class of similar historical examples because that is a problem of general intelligence. Once identified, however, we want to rely on ML approaches because, with the small number of examples at hand compared to the amount of information we have about each of them, we are facing a problem of high-dimensional data, where ML outperforms.
Viewed this way, as a technique that incorporates the abilities of both humans and ML, each where they perform best, it makes sense that superforecasting is able to make more accurate predictions than any other approach.
