Estimating value at risk: LSTM vs. GARCH

TL;DR

This paper introduces a novel LSTM-based estimator for value-at-risk that outperforms traditional GARCH models on real market data, especially in capturing volatility dynamics.

Contribution

The paper presents the first application of LSTM neural networks for value-at-risk estimation, demonstrating superior performance on market data compared to GARCH models.

Findings

LSTM performs similarly to GARCH on simulated data.

LSTM better captures volatility shifts on real market data.

LSTM outperforms existing estimators in exception rate and quantile score.

Abstract

Estimating value-at-risk on time series data with possibly heteroscedastic dynamics is a highly challenging task. Typically, we face a small data problem in combination with a high degree of non-linearity, causing difficulties for both classical and machine-learning estimation algorithms. In this paper, we propose a novel value-at-risk estimator using a long short-term memory (LSTM) neural network and compare its performance to benchmark GARCH estimators. Our results indicate that even for a relatively short time series, the LSTM could be used to refine or monitor risk estimation processes and correctly identify the underlying risk dynamics in a non-parametric fashion. We evaluate the estimator on both simulated and market data with a focus on heteroscedasticity, finding that LSTM exhibits a similar performance to GARCH estimators on simulated data, whereas on real market data it is more sensitive towards increasing or decreasing volatility and outperforms all existing estimators of value-at-risk in terms of exception rate and mean quantile score.