KILLED · overfit by design
Hypothesis
Evolving populations of indicator-threshold rules with a genetic algorithm discovers profitable combinations no human would hand-craft.
Math — selection pressure
Fitness-proportional selection over generations:
$$ P(\text{rule}_i) = \frac{\phi_i}{\sum_j \phi_j},\quad \phi=\text{in-sample Sharpe} $$
Method
GA over indicator/threshold/logic genes, fitness = in-sample Sharpe, 200 generations, then hold-out test.
Results
| Best in-sample Sharpe | 4.1 |
| Same rule out-of-sample | −0.3 |
| Multiple-testing inflation | extreme |
Evolving toward in-sample Sharpe is industrial-scale data snooping: searching millions of rule combinations guarantees a spectacular in-sample winner that is pure overfit. Out-of-sample it is negative. The GA optimizes the one thing you must not optimize directly.
If your search procedure maximizes in-sample performance over a huge rule space, your "discovery" is the maximum of noise. Significance must be penalized by the number of trials.