SciML in the Wild: A Diagnostic Study of When Structural Priors Help and When They Hurt

Short summary

This study evaluates five SciML model families (ARIMA, LSTM, NODE, PINN, UDE) on macroeconomic forecasting across 23 countries, finding that less-constrained models consistently outperform more-constrained heuristic-prior models. Structural priors act as misregularizers when they don't match the data-generating process. The authors identify failure modes including prior misalignment, regime shifts, and optimization instability, recommending practitioners test whether structure helps before assuming it does.

• •Less-constrained models (ARIMA, NODE) outperform constrained prior models (PINN, UDE) in macroeconomic forecasting
• •Structural priors can act as misregularizers when they don't match the data-generating process
• •Practitioners should empirically test whether structure helps before assuming more structure is beneficial

Generated with AI, which can make mistakes.