TSFM_VaR_CEE

Reproduction code for the paper:

Can Foundation Models Manage Risk? Zero-Shot VaR and ES Forecasting with Conformal Calibration in CEE Markets

Siang-Li Jheng, Rahul Tak, Ștefan Găman, Miruna Mazurencu-Marinescu-Pele, Daniel Traian Pele (corresponding author — danpele@ase.ro)

Bucharest University of Economic Studies · Institute for Economic Forecasting, Romanian Academy

Economic Computation and Economic Cybernetics Studies and Research (ECECSR), 2026.

Overview

Zero-shot Value-at-Risk (VaR) and Expected Shortfall (ES) forecasting for five CEE equity indices (BET, WIG20, PX, BUX, SOFIX) and five FX pairs (EUR/RON, EUR/PLN, EUR/CZK, EUR/HUF, USD/BGN), comparing time-series foundation models (Chronos-2, TimesFM 2.5, Moirai 2.0) — raw and conformally calibrated — against econometric baselines (GJR-GARCH, Historical Simulation, ARIMA-Conformal, LSTM-Conformal) under the Basel traffic-light framework.

Each Quantlet is a self-contained Jupyter notebook that imports shared settings from config.py.

Repository structure (Quantlets)

Quantlet	Purpose
VaR_CEE_DataDownload	Download raw index / FX series
VaR_CEE_DataPipeline	Clean, align, compute log-returns
VaR_CEE_GJRGARCH	GJR-GARCH baseline VaR / ES
VaR_CEE_HistoricalSim	Historical Simulation baseline
VaR_CEE_Chronos	Chronos-2 zero-shot forecasts
VaR_CEE_TimesFM	TimesFM zero-shot forecasts
VaR_CEE_Moirai	Moirai zero-shot forecasts
VaR_CEE_ConformalARIMA	ARIMA + conformal prediction
VaR_CEE_ConformalLSTM	LSTM + conformal prediction
VaR_CEE_ConformalFM	Conformal calibration of TSFMs
VaR_CEE_Backtesting	Kupiec, Christoffersen, Acerbi-Szekely, Basel zones
VaR_CEE_DMTest	Diebold-Mariano tests
VaR_CEE_Figures	All paper figures

config.py — shared paths, market definitions, and parameters used by every Quantlet (sample 2007-01-01 to 2025-12-31; out-of-sample from 2018-01-01; rolling window 250 days; VaR levels 1%, 2.5%, 5%; ES level 2.5%; foundation-model context 512, 1000 samples, 1-step horizon; seed 42).

How to run

Edit config.py if needed, then execute the notebooks in this order:

jupyter nbconvert --to notebook --execute VaR_CEE_DataDownload/VaR_CEE_DataDownload.ipynb
jupyter nbconvert --to notebook --execute VaR_CEE_DataPipeline/VaR_CEE_DataPipeline.ipynb
jupyter nbconvert --to notebook --execute VaR_CEE_GJRGARCH/VaR_CEE_GJRGARCH.ipynb
jupyter nbconvert --to notebook --execute VaR_CEE_HistoricalSim/VaR_CEE_HistoricalSim.ipynb
jupyter nbconvert --to notebook --execute VaR_CEE_Chronos/VaR_CEE_Chronos.ipynb
jupyter nbconvert --to notebook --execute VaR_CEE_TimesFM/VaR_CEE_TimesFM.ipynb
jupyter nbconvert --to notebook --execute VaR_CEE_Moirai/VaR_CEE_Moirai.ipynb
jupyter nbconvert --to notebook --execute VaR_CEE_ConformalARIMA/VaR_CEE_ConformalARIMA.ipynb
jupyter nbconvert --to notebook --execute VaR_CEE_ConformalLSTM/VaR_CEE_ConformalLSTM.ipynb
jupyter nbconvert --to notebook --execute VaR_CEE_ConformalFM/VaR_CEE_ConformalFM.ipynb
jupyter nbconvert --to notebook --execute VaR_CEE_Backtesting/VaR_CEE_Backtesting.ipynb
jupyter nbconvert --to notebook --execute VaR_CEE_DMTest/VaR_CEE_DMTest.ipynb
jupyter nbconvert --to notebook --execute VaR_CEE_Figures/VaR_CEE_Figures.ipynb

Outputs are written to data/ (raw, processed, var_forecasts, backtesting, figures, stats).

Environment

Tested with Python 3.13. The three foundation models need separate virtual environments because their dependencies conflict; the baselines, conformal methods, backtesting and figures share a common environment.

Environment	Install	Used by
Common	pip install -r requirements.txt	DataDownload, DataPipeline, GJRGARCH, HistoricalSim, ConformalARIMA, ConformalLSTM, Backtesting, DMTest, Figures
Chronos	pip install -r requirements-chronos.txt	Chronos (+ ConformalFM)
TimesFM	pip install -r requirements-timesfm.txt	TimesFM (+ ConformalFM)
Moirai	pip install -r requirements-moirai.txt	Moirai (+ ConformalFM)

Data

Daily equity-index and FX series from Stooq / Yahoo Finance (see VaR_CEE_DataDownload for tickers and retrieval).

Citation

See CITATION.cff.

License

MIT — see LICENSE.