Combining dimensionality reduction with neural networks for realized volatility forecasting

Abstract

The application of artificial neural networks to finance has received a great deal of attention from both investors and researchers, especially as a forecasting method. When the number of predictors is high, these methods suffer from the so-called "curse of dimensionality" and produce biased forecasts. In this paper, we relied on dimensionality reduction methods to alleviate such issue when a wide set of financial and macroeconomic variables is considered in the prediction of stock market volatility. Specifically, we combined Bayesian Model Averaging (BMA), Principal Component Analysis (PCA), Non-negative Matrix Factorization (NMF) and Least Absolute Shrinkage and Selection Operator (LASSO) with hybrid artificial neutral networks to forecast realized volatility. The results showed that reduced models were able to perform in a similar way or even outperforms the compared full models in terms of predictive accuracy.