Title: A machine learning approach to volatility forecasting
Authors: Mathias Siggaard - Aarhus University (Denmark) [presenting]
Abstract: This paper shows how machine learning algorithms can improve the forecast accuracy of one-day-ahead forecast for high-frequency for volatility series. To achieve this, all stocks from Dow Jones Industrial Average index over the sample period from 2001 to 2018 are examined. Four groups of machine learning algorithms are compared; Regularization Methods, Tree-Based Methods, Deep Learning, and Ensemble Methods. Comparison with the commonly-used Heterogeneous Autoregressive model shows substantial improvement. Through model confidence set and forecast comparison, the best machine learning algorithms (neural networks and random forest) are identified. Furthermore, it is shown how these methods are capable of extracting important information when including additional explanatory variables and find a small set of dominating predictors including implied volatility, earnings announcements, and a daily policy index for United States. To better understand how these methods behave under extreme situations, this paper investigates their performance during the flash crash of 24th of August 2015.