Quantitative Overfitting
Quantitative overfitting is the danger that a trading rule looks profitable in historical data but fails outside the sample. In EP88 穿越量化之父西蒙斯：AI会让普通人更容易赚钱，还是更难？, Jim Simons’s episode persona treats it as one of the main reasons ordinary investors should distrust backtests, seasonal patterns, and widely circulated “holy grail” strategies.
E144.交易的艺术：不预测，统计优势，分散红利，随机波动 adds Random Market Narratives as the narrative counterpart. Its random “山海经神兽” experiment shows how convincing explanations can be generated from price paths even when the underlying process was designed to have no real causal story.
Key Claims
- A pattern needs a plausible mechanism, not only an attractive chart.
- Out-of-sample testing matters because history contains many accidental regularities.
- Simpler and more robust rules are preferred over complicated rules that perfectly fit past noise.
- Machine learning does not excuse blind pattern mining; researchers still need hypotheses and judgment.
- Overfit signals can decay quickly once they are public or once market conditions change.
- A random experiment can still produce apparent winners, sectors, and themes, so backtests need humility about causality.
Connections
- Quantitative Investing — method most exposed to this statistical failure mode.
- Market Regime Shift — another reason a historical strategy can stop working.
- Market Efficiency — alpha is small and temporary when markets are close to efficient.
- AI Investment Research — AI can accelerate both useful analysis and spurious pattern discovery.
- Random Market Narratives and Behavioral Investing Biases — E144’s story-generation warning around random price paths.