[2603.02620] Same Error, Different Function: The Optimizer as an Implicit Prior in Financial Time Series

Computer Science > Machine Learning arXiv:2603.02620 (cs) [Submitted on 3 Mar 2026] Title:Same Error, Different Function: The Optimizer as an Implicit Prior in Financial Time Series Authors:Federico Vittorio Cortesi, Giuseppe Iannone, Giulia Crippa, Tomaso Poggio, Pierfrancesco Beneventano View a PDF of the paper titled Same Error, Different Function: The Optimizer as an Implicit Prior in Financial Time Series, by Federico Vittorio Cortesi and 4 other authors View PDF HTML (experimental) Abstract:Neural networks applied to financial time series operate in a regime of underspecification, where model predictors achieve indistinguishable out-of-sample error. Using large-scale volatility forecasting for S$\&$P 500 stocks, we show that different model-training-pipeline pairs with identical test loss learn qualitatively different functions. Across architectures, predictive accuracy remains unchanged, yet optimizer choice reshapes non-linear response profiles and temporal dependence differently. These divergences have material consequences for decisions: volatility-ranked portfolios trace a near-vertical Sharpe-turnover frontier, with nearly $3\times$ turnover dispersion at comparable Sharpe ratios. We conclude that in underspecified settings, optimization acts as a consequential source of inductive bias, thus model evaluation should extend beyond scalar loss to encompass functional and decision-level implications. Comments: Subjects: Machine Learning (cs.LG); Computational Finan...