The Forecasting Grand Prix: Self-Driving Neural Networks for Inflation Prediction

Authors: Sicco Kooiker, Janneke van Brummelen, and Julia Schaumburg

Abstract:

We propose a time-varying neural network framework for macroeconomic forecasting in which the weights and biases of a neural network evolve over time using a self-driving up- dating scheme based on past forecast errors. While maintaining the flexibility of neural networks, we introduce observation-driven adaptation through parameter dynamics, linking the machine learning and observation-driven parameter literatures. To address sensitivity to hyperparameter choices, a stacked recursive least squares method find the best performing hyperparameters on-the-fly. The method is fully online and does not require re-estimation over a rolling window. Using predictors from FRED-MD, we evaluate multi-horizon U.S. inflation forecasts over the most recent 15 years against a broad set of machine learning benchmarks. We show that the proposed method improves forecast accuracy relative to static neural network specifications and competing benchmarks. In-sample diagnostics illus- trate how the optimal hyperparameter configuration and variable importance change over time.

Presentations:

• ISF Conference, Montreal, Canada (2026)

Recommended citation: Kooiker, S., van Brummelen, J. & Schaumburg, J. (2026). The Forecasting Grand Prix: Self-Driving Neural Networks for Inflation Prediction. Working Paper.
Download Paper