Overfitting in Human Decision-Making

The human tendency to build overly complex explanations that perfectly fit past data but fail to predict future events—finding patterns in noise and building theories around them. A manager who develops a detailed theory about why the last three hires succeeded, then hires based on that theory and fails—because the 'pattern' was noise in a small sample. An investor who identifies a 'formula' from five successful trades that fails on the sixth—because the formula captured randomness, not signal. Overfitting is the human version of the machine learning problem: with enough parameters, you can explain anything retrospectively but predict nothing prospectively.

When to use it

When complex theories are built from small samples; when past patterns don't predict future outcomes; when explanations seem to fit historical data perfectly but fail prospectively; when the bias-variance tradeoff needs to be managed in human judgment.

How it can help

Favor simple models over complex ones (Occam's Razor applied to decision-making). The diagnostic: does your explanation have as many moving parts as data points? If so, you've probably overfit. The corrections: (1) Demand out-of-sample validation—does the pattern hold in NEW data, not just the data it was built on? (2) Prefer models with fewer parameters. (3) Be especially suspicious of patterns found in small samples. (4) Use cross-validation—test your theory on different time periods or contexts. The rule: the more complex your explanation, the more likely it captures noise rather than signal.