Best model from an evoloop evolutionary optimization campaign.

Experiment: exp_heston_garch_dynamic_df_scaling (index 223, generation 17)

Motivation: The research program has stalled at 0.950661 GBM-normalized CRPS with 30+ consecutive convergence confirmations. The "Uncertainty & Horizon Scaling" strategy category remains untried. I will implement dynamic, time-varying Student-t degrees of freedom that scale inversely with recent realized volatility—heavier tails (lower df) during high-vol periods, lighter tails (higher df) during calm periods—to better capture regime-dependent tail behavior without adding per-asset calibration parameters. Hypothesis: Static df parameters force a compromise across volatility regimes. By dynamically adjusting df based on recent realized volatility (df_t = base_df * (1 + vol_ratio)), the model can adapt tail heaviness to current market conditions. This should improve CRPS during regime transitions (where the static model under-predicts tail risk) while maintaining calibration in stable periods. The mechanism leverages the "volatility of volatility" signal already present in the Heston component to drive distributional flexibility without structural overfitting.

Reflection: ## Reflection

This experiment teaches us that the statistical ceiling is an information-theoretic boundary, not a model-specification boundary. Even when we introduce a theoretically-sound mechanism from an untried category (dynamic df scaling based on volatility regimes), the effect size remains trapped at 0.016% of SE—indistinguishable from noise. The Heston+GARCH hybrid already extracts all available predictive signal from 5-minute price data for 24h horizons; remaining "improvements" are just reshaping noise. This validates that the 30+ convergence streak represented true epistemic closure: the asymptotic limit is fundamental to the data generating process, not an artifact of model family limitations. The scientifically appropriate action is definitive termination—pursuing sub-noise effects has negative expected value regardless of theoretical motivation.

This model was evolved using evoloop, an LLM-driven evolutionary code optimization system. Starting from a seed baseline, the system iteratively mutated the training code across 252 experiments over 19 generations, guided by CRPS scoring and research hypotheses.