Better prediction error but worse net returns

A new model lowers return-prediction MSE but produces lower net strategy returns. Design an investigation.

Hold the evaluation sample and execution assumptions fixed. Compare improvements by trade size, horizon and decision threshold; the model may improve untraded observations while harming selected trades. Inspect turnover, spreads, impact, risk exposure and calibration at the traded tail. Separate forecast quality from portfolio construction and costs. Tune any revised decision rule only within development periods and evaluate it on an untouched outer sample.