Long Short-Term Memory neural network for econometric forecasting: A comparison between a statistical method and a neural network in the case of Value at Risk

Details

Supervisors

Hainaut, Donatien

Faculty

Faculté des sciences

Degree label

Master [120] en science des données, orientation statistique, à finalité spécialisée

Abstract

The objective of this master thesis is to provide a comprehensive comparison between a benchmark econometric model and a Long Short-Term Memory Neural Network on the computation of the Value at Risk of the Standard & Poor’s Index (S&P500). To do so, we model the conditional mean and volatility of the financial series using these two multivariate models and proceed to Monte Carlo simulations to retrieve the necessary quantile. For the benchmark econometric model, we use a Vector AutoRegressive Moving Average to estimate the conditional mean. To estimate the conditional volatility, we adopted the Dynamic Conditional Correlation model. Regarding the neural network, the architecture is built with two blocks. The first one is used for feature extraction purpose and the second one is used to forecast the parameters needed based on the latent codings and the exogenous variables. From the one-day ahead and 10-days ahead conditional estimations, we simulate 10.000 scenarios of returns and compute the Value at Risk and Expected Shortfall. We conclude this master thesis by comparing the performances on both risk measure exceedances and suggest ways to improve the computations.