PREDICTING IBOVESPA RISK USING A HYBRID AESM-MLCP APPROACH

Abstract

This study proposes and evaluates a hybrid model for forecasting the daily risk of the IBOVESPA, combining Multivariate Singular Spectrum Analysis (MSSA) and Long Short-Term Memory (LSTM) neural networks. The research is characterized as applied, quantitative, explanatory, documental, computational, and empirical-analytical, using daily data from the IBOVESPA, IBrX-100, and IBrX-50 indices, obtained from the ADVFN portal, covering the period from February 15, 2021, to March 27, 2026. The risk variable was constructed using the natural logarithm of the ratio between the daily high and low prices. Methodologically, the study was structured into two stages: in the first, an automatic optimization of MSSA and LSTM parameters was performed; in the second, the final training was conducted using the best configuration identified. MSSA was employed as a structural filtering technique and for extracting common components across the series, while LSTM was used as a predictive model to capture nonlinear temporal dependencies. The results demonstrated superior performance of the hybrid approach compared to preliminary strategies, particularly in terms of adherence between the predicted and observed series. The best configuration identified during training was r=5, L=10 and lookback = 30, achieving a correlation of 0.860537 in the test set, along with a mean absolute error of 0.051964 and a root mean squared error of 0.066677, calculated on the normalized variable. In the testing phase, correlation reached 0.90 with lower error measures.It is concluded that the integration of MSSA and LSTM constitutes a promising alternative for risk forecasting in financial time series, although limitations remain in fully capturing extreme events and in the absence of robust temporal validation.