Value-at-Risk prediction for the Brazilian stock market: A comparative study between Parametric Method, Feedforward and LSTM Neural Network

Abstract

Value-at-Risk (VaR) as a risk quantification mechanism has more than one calculation method, one of which is the parametric method. In the bibliographic study, it was identified that the parametric method is not effective for all market moments, such as those of crisis or abrupt changes in behavior. This study, therefore, seeks to verify whether other methods of calculation are more efficient, such as the use of neural networks. This study compared the VaR calculation using the parametric method against the use of Feedforward neural networks and Long Short-Term Memory (LSTM) recurrent networks. It used the B3 Sao Paulo Stock Exchange, IBOVESPA, as the index of study. For the parametric method, volatility models such as standard deviation, Exponentially Weighted Moving Average (EWMA) and Generalized Autoregressive Conditional Heteroskedasticity (GARCH) were tested. For neural networks, different layers and amounts of neurons, activation functions, use of different predictors and incorporation of macroeconomic data were explored. The result of the experiment showed that LSTM networks had a better performance when comparing the exception rate generated by the entire model. When analyzing periods of crisis or abrupt changes in behavior, LSTM and Feedforward networks were less efficient in predicting VaR compared to the parametric method.