SciML Study Shows Structural Priors Can Act as Misregularizers in Forecasting

A new diagnostic study from Vrishank Sai Anand, Prathamesh Dinesh Joshi, Raj Abhijit Dandekar, Rajat Dandekar, and Sreedath Panat tests when structural priors in Scientific Machine Learning (SciML) help or hurt. The researchers evaluate five model families—ARIMA, LSTM, Neural ODEs (NODEs), Physics-Informed Neural Networks (PINNs), and Universal Differential Equations (UDEs)—across 23 countries using sparse annual data with multiple temporal splits and random seeds. The results are stark: none of the models achieve consistently strong forecasting performance, highlighting the difficulty of low-frequency macroeconomic prediction. However, a clear hierarchy emerges—less-constrained models (ARIMA and NODE) consistently outperform more-constrained heuristic-prior models (PINN and UDE).

The authors interpret this as a diagnostic result: structural priors act as misregularizers when they do not match the data-generating process. They identify specific failure modes including prior misalignment, regime shifts, structural breaks, and optimization instability. The paper argues that SciML practitioners should test whether structure helps before assuming more structure is beneficial. This marks a critical caution for the field, especially in domains like economics where reliable governing dynamics are rare. The study is available on arXiv (2607.09684) and serves as a practical guide for when to lean on structural priors versus letting the data speak.