A1459
Title: Macroeconomic forecasting with time series foundation models
Authors: Rolf Scheufele - Swiss National Bank (Switzerland) [presenting]
Abstract: Time series foundation models (TSFMs) such as TimesFM and Chronos promise major advances in forecasting, yet their evaluation is hampered by limited control over training data and the infeasibility of standard pseudo out-of-sample tests. The purpose is to address this by conducting a real-time forecasting experiment using daily financial and macroeconomic series (exchange rates, interest rates, stock prices) from June 2025 onward. This high-frequency design yields genuine out-of-sample observations, allowing a fair comparison between leading TSFMs and traditional benchmarks. Results show heterogeneous performance: while some TSFMs match or slightly outperform univariate models and the random walk, others fall short. Early evidence is provided on the opportunities and limitations of TSFMs in economics and finance.
