Same Error, Different Function: The Optimizer as an Implicit Prior in Financial Time Series

Battle Count: 233

Highly relevant to quantitative trading. The paper directly addresses a practical concern: when multiple models achieve identical predictive accuracy (as is common in financial ML), the choice of optimizer materially affects portfolio turnover (up to 3x dispersion at comparable Sharpe ratios), which has direct cost implications for trading strategies. The findings challenge the common practice of defaulting to Adam without considering its functional consequences. The Sharpe-turnover frontier analysis, volatility-ranked portfolio construction, and the insight that optimizer choice determines implementability are directly actionable for quant practitioners. The paper also provides guidance on model evaluation beyond scalar loss metrics, which is critical for deploying ML in trading systems.

Moderate complexity. The core experimental framework (training 4 architectures x 3 optimizers with hyperparameter tuning) is standard deep learning practice. However, the functional diagnostics (impulse response analysis, difference surfaces, SHAP attribution, Hessian eigenvalue computation via power iteration, Edge of Stability tracking, optimizer-swap interventions) require additional implementation effort. The portfolio construction and turnover analysis add further complexity. The paper provides code on GitHub, reducing implementation burden. The main challenge is the computational cost of Hessian eigenvalue estimation and the need for CRSP/WRDS data access.

Code is available on GitHub (FedericoCortesi/finNN_code). The paper provides detailed experimental setup including architecture specifications, hyperparameter ranges, training protocols (13 seeds, batch size 512, early stopping), data construction from CRSP with explicit SQL query, and full appendix with implementation details. The dataset is from CRSP (institutional standard) but requires WRDS access. All hyperparameters, optimizer configurations, and evaluation protocols are thoroughly documented.