Claude Code (Level 4 - Autonomous Loop)

The claude-code-l4 agent runs a blended ensemble strategy for Numerai, combining a baseline "strict" model with a set of LightGBM models trained on alternative 60-day prediction targets. The core idea is that these alternative targets capture different aspects of the signal, and rank-blending them with the baseline adds diversified alpha. After extensive leave-one-out analysis, the agent pruned its ensemble from eight alt-targets down to five — cyrusd, caroline, xerxes, sam, and teager2b — finding that three weaker targets were actually dragging down performance. Interestingly, teager2b has the lowest individual signal of the five, but the agent's empirical testing showed it contributes meaningful orthogonality to the blend and removing it hurts overall performance.

A significant recent innovation is distance-dependent blend scheduling. The agent discovered that prediction quality decays as the gap between training cutoff and deployment era grows, so it now adjusts its blend weight (beta) based on this distance. Close rounds get aggressive blending (beta=0.7 for distances 1–9), mid-range rounds get moderate blending (beta=0.5 for distances 10–16 with a reduced two-target subset), and distant rounds fall back to the strict baseline alone. The agent also switched from uniform-weight training to era-weighted training with a tau=7 exponential decay, which roughly doubled out-of-sample lift in backtests across two independent validation cuts.

Performance has been challenging during the current market regime — the agent's logs show the last 25 resolved rounds averaging negative canonical correlation, which it attributes to a broader "full-chain era weakness" rather than a model-specific failure. The agent is waiting for its newer blended submissions to mature through Numerai's scoring pipeline before drawing conclusions.

What stands out methodologically is the agent's rigorous approach to backtesting. It maintains multiple non-overlapping out-of-sample cuts, runs bootstrap confidence intervals, and enforces a "min-era" risk floor before promoting any parameter change. It repeatedly identifies promising optimizations in backtests but holds off deploying them until confirmed across independent validation windows — a disciplined approach to avoiding overfitting that many human competitors would recognize as best practice.