A1472
Title: Probabilistic AI for improved tail risk estimation
Authors: Morten Risstad - NTNU (Norway) [presenting]
Rickard Sandberg - Stockholm School of Economics (Sweden)
Abstract: Probabilistic AI models are explored for day-ahead forecasting of return distributions and tail risk. Traditional econometric models rely on restrictive parametric assumptions, while standard machine learning offers only point forecasts without uncertainty. Probabilistic AI addresses these gaps by directly generating flexible conditional return distributions with uncertainty quantification. LSTM and transformer architectures are implemented within a mixture density network (MDN) framework, using realized variance from the CaPiRe dataset and implied volatility from Bloomberg as predictors. The models yield non-parametric return distributions for DJIA constituents, from which value-at-risk (VaR) and expected shortfall (ES) are derived. Performance is compared against econometric and machine learning benchmarks using statistical adequacy tests and scoring metrics. Predictive uncertainty is further decomposed into aleatoric (market risk) and epistemic (model risk) components, enhancing interpretability. Empirical results show that LSTM-MDN and Transformer-MDN consistently outperform benchmarks in distributional accuracy, calibration, and tail risk forecasting, remaining robust across market regimes, including COVID-19. Models incorporating implied volatility deliver the strongest performance, confirming their superior predictive power. Probabilistic AI thus offers a powerful alternative for risk management, portfolio allocation, and derivative pricing.
