Do Artificial Neural Networks Provide Improved Volatility Forecasts: Evidence from Asian Markets

Abstract

This paper enters the ongoing volatility forecasting debate by examining the ability of a wide range of Machine Learning methods (ML), and specifically Artificial Neural Network (ANN) models. The ANN models are compared against traditional econometric models for ten Asian markets using daily data for the time period from 12 September 1994 to 05 March 2018. The empirical results indicate that ML algorithms, across the range of countries, can better approximate dependencies compared to traditional benchmark models. Notably, the predictive performance of such deep learning models is superior perhaps due to its ability in capturing long-range dependencies. For example, the Neuro Fuzzy models of ANFIS and CANFIS, which outperform the EGARCH model, are more flexible in modelling both asymmetry and long memory properties. This offers new insights for Asian markets. In addition to standard statistics forecast metrics, we also consider risk management measures including the value-at-risk (VaR) average failure rate, the Kupiec LR test, the Christoffersen independence test, the expected shortfall (ES) and the dynamic quantile test. The study concludes that ML algorithms provide improving volatility forecasts in the stock markets of Asia and suggest that this may be a fruitful approach for risk management.