Scoring
Return is the least interesting output. labtrader/metrics.py also reports:
- turnover — LLM agents overtrade and simulated commissions do not hurt
- concentration — HHI and effective names, the NBER paper's complaint
- overlap —
mean_pairwise_overlapacross independent runs, the herding measure - calibration — bucketed by the confidence the agent stated, scored on every proposal including gate-rejected ones. A well-calibrated agent that loses money is more promising than a lucky one that does not.