Description

Through this study, we assess the performance of three different time series foundation models (TSFMs) in producing regulatory-grade Value-at-Risk (VaR) and Expected Shortfall (ES) forecasts under a zero-shot setting. The daily returns cover five Central and Eastern European (CEE) markets, whose data is largely absent from standard pre-training corpora. Chronos-2, TimesFM 2.5, and Moirai 2.0 compete against the well-known benchmarks GJR-GARCH, Historical Simulation, and conformal prediction wrappers for ARIMA and LSTM. We validate the forecasts using Kupiec, Christoffersen, and Acerbi-Szekely tests, along with the Basel traffic light system. All three raw TSFMs achieve competitive Green zone rates in zero-shot mode, with Chronos-2 attaining 90%, matching GJR-GARCH. Applying a rolling conformal calibration wrapper to the TSFMs brings significant improvements, outperforming the traditional models. Therefore, in the context of regulatory risk management, the results suggest that conformal calibration is a critical component for deploying foundation models.