Moderating Intuitive Predictions and Regression to the Mean

A DEFENSE OF EXTREME PREDICTIONS?

Predictions of discrete outcomes, such as a student’s field of specialization or success in an examination, assign probabilities to events or rank outcomes from most to least probable. They are vulnerable to neglect of base rates and insensitivity to the quality of information.

The biases we find in predictions that are expressed on a scale, such as GPA or the revenue of a firm, are similar to the biases observed in judging the probabilities of outcomes.

The corrective procedures are also similar:

• Both contain a baseline prediction, which you would make if you knew nothing about the case at hand. In the categorical case, it was the base rate. In the numerical case, it is the average outcome in the relevant category.

• Both contain an intuitive prediction, which expresses the number that comes to your mind, whether it is a probability or a GPA.

• In both cases, you aim for a prediction that is intermediate between the baseline and your intuitive response.

• In the default case of no useful evidence, you stay with the baseline.

• At the other extreme, you also stay with your initial prediction. This will happen, of course, only if you remain completely confident in your initial prediction after a critical review of the evidence that supports it.

• In most cases you will find some reason to doubt that the correlation between your intuitive judgment and the truth is perfect, and you will end up somewhere between the two poles.

This procedure is an approximation of the likely results of an appropriate statistical analysis. If successful, it will move you toward unbiased predictions, reasonable assessments of probability, and moderate predictions of numerical outcomes. The two procedures are intended to address the same bias: intuitive predictions tend to be overconfident and overly extreme.

Correcting your intuitive predictions is a task for System 2. Significant effort is required to find the relevant reference category, estimate the baseline prediction, and evaluate the quality of the evidence. The effort is justified only when the stakes are high and when you are particularly keen not to make mistakes. Furthermore, you should know that correcting your intuitions may complicate your life. A characteristic of unbiased predictions is that they permit the prediction of rare or extreme events only when the information is very good. If you expect your predictions to be of modest validity, you will never guess an outcome that is either rare or far from the mean. If your predictions are unbiased, you will never have the satisfying experience of correctly calling an extreme case. You will never be able to say, “I thought so!” when your best student in law school becomes a Supreme Court justice, or when a start-up that you thought very promising eventually becomes a major commercial success. Given the limitations of the evidence, you will never predict that an outstanding high school student will be a straight-A student at Princeton. For the same reason, a venture capitalist will never be told that the probability of success for a start-up in its early stages is “very high.”

The objections to the principle of moderating intuitive predictions must be taken seriously, because absence of bias is not always what matters most. A preference for unbiased predictions is justified if all errors of prediction are treated alike, regardless of their direction. But there are situations in which one type of error is much worse than another. When a venture capitalist looks for “the next big thing,” the risk of missing the next Google or Facebook is far more important than the risk of making a modest investment in a start-up that ultimately fails. The goal of venture capitalists is to call the extreme cases correctly, even at the cost of overestimating the prospects of many other ventures. For a conservative banker making large loans, the risk of a single borrower going bankrupt may outweigh the risk of turning down several would-be clients who would fulfill their obligations. In such cases, the use of extreme language (“very good prospect,” “serious risk of default”) may have some justification for the comfort it provides, even if the information on which these judgments are based is of only modest validity.

For a rational person, predictions that are unbiased and moderate should not present a problem. After all, the rational venture capitalist knows that even the most promising start-ups have only a moderate chance of success. She views her job as picking the most promising bets from the bets that are available and does not feel the need to delude herself about the prospects of a start-up in which she plans to invest. Similarly, rational individuals predicting the revenue of a firm will not be bound to a single number—they should consider the range of uncertainty around the most likely outcome. A rational person will invest a large sum in an enterprise that is most likely to fail if the rewards of success are large enough, without deluding herself about the chances of success. However, we are not all rational, and some of us may need the security of distorted estimates to avoid paralysis. If you choose to delude yourself by accepting extreme predictions, however, you will do well to remain aware of your self-indulgence.

Perhaps the most valuable contribution of the corrective procedures I propose is that they will require you to think about how much you know. I will use an example that is familiar in the academic world, but the analogies to other spheres of life are immediate. A department is about to hire a young professor and wants to choose the one whose prospects for scientific productivity are the best. The search committee has narrowed down the choice to two candidates:

Kim recently completed her graduate work. Her recommendations are spectacular and she gave a brilliant talk and impressed everyone in her interviews. She has no substantial track record of scientific productivity.

Jane has held a postdoctoral position for the last three years. She has been very productive and her research record is excellent, but her talk and interviews were less sparkling than Kim’s.

The intuitive choice favors Kim because she left a stronger impression. This is WYSIATI—what you see is all there is: the vivid interview dominates the information that is missing. But it is also the case that there is much less information about Kim than about Jane. We are back to the law of small numbers. In effect, you have a smaller sample of information from Kim than from Jane, and extreme outcomes are much more likely to be observed in small samples. There is more luck in the outcomes of small samples, and you should therefore regress your prediction more deeply toward the mean in your prediction of Kim’s future performance. When you allow for the fact that Kim is likely to regress more than Jane, you might end up selecting Jane although you were less impressed by her. In the context of academic choices, I would vote for Jane, but it would be a struggle to overcome my intuitive impression that Kim is more promising. Following our intuitions is more natural, and somehow more pleasant, than acting against them.

You can readily imagine similar problems in different contexts, such as a venture capitalist choosing between investments in two start-ups that operate in different markets. One start-up has a product for which demand can be estimated with fair precision. The other candidate is more exciting and intuitively promising, but its prospects are less certain. Whether the best guess about the prospects of the second start-up is still superior when the uncertainty is factored in is a question that deserves careful consideration.