Sistemas híbridos para previsão de volatilidade em investimentos: GARCH e MLP no mercado brasileiro

Abstract

Over recent decades, Brazil’s capital market has seen growing retail participation and recurrent volatility episodes, heightening the demand for more accurate market-risk forecasts. To address linear and nonlinear dynamics in financial time series, this study proposes a hybrid model that integrates GARCH—which captures conditional heteroskedasticity—with an MLP neural network aimed at nonlinear patterns. We use daily data for Ibovespa (IBOV), Brazil 100 (IBRX100), Dividend Index (DIVI), Corporate Sustainability Index (ISEE), and Small Cap (SMLL) from 2017–2023, computing log returns and estimating conditional volatility. Accuracy is assessed via MSE, MAE, and QLIKE. Results show MSE ≈ 0 and that GARCH achieves the lowest MAE for most indices (IBOV, IBRX100, ISEE, SMLL), while MLP reduces MAE only for DIVI (~+3.09% vs. GARCH). For QLIKE, the hybrid model attains the lowest values across all indices, indicating better coverage of volatility spikes; MLP also improves QLIKE for four series (with a slight deterioration only for DIVI). These findings support a complementary use: GARCH as the scale-calibration anchor with lower typical error in most cases, and MLP/Hybrid for better tail and spike coverage (lower QLIKE) with higher responsiveness. Within the scope of univariate volatility series, this choice aligns with the study’s objectives. Limitations include using a single nonlinear neural network architecture (MLP) and no exogenous variables. Future work can explore model-combination learners,

dynamic predictor selection, and complementary deep-learning architectures (e.g., LSTM, N- BEATS, Transformers), while enriching inputs with macro and sentiment information to

enhance robustness and practical applicability in dynamic markets.