Market volatility

: can machine learning methods enhance volatility forecasting?

Abstract

This dissertation aims to test whether the use of machine learning (ML) techniques can improvevolatility forecasting accuracy. More specifically, if it can beat the best econometric model, theHeterogeneous Autoregressive model of Realized Volatility (HAR-RV). Using S&P 500 Indexdata from May-2007 to August-2022, the superiority of the HAR-RV was tested and attestedagainst competing econometric models EWMA and GARCH(1,1). Next, the performance ofthe ML Artificial Neural Network algorithms Long Short-Term Memory (LSTM) and GatedRecurrent Unit (GRU) are compared to the performance of the econometric models. Fivedifferent variable sets are tested for the ML models. It is found that while both ML models areable to beat the EWMA and GARCH(1,1) models by a significant margin, the HAR-RV modelstill outperforms LSTM and GRU.Moreover, an analysis is conduced on the models’ predictions on the period corresponding tothe Covid-19 crisis. The results did not show any evidence suggesting that ML methods havea particular advantage at predicting during high volatility events. Finally, a plausible cause that could undermine the remarkable qualities of the ML methods inthe aim of volatility forecasting is discussed. It is found that the rigorous set of conditionsneeded to be met for the proper setup of ML models are very difficult to be met using financialdata, which hinders the aptitude of ML for this purpose.

Date of Award	5 May 2023
Original language	English
Awarding Institution	Universidade Católica Portuguesa
Supervisor	José Faias (Supervisor)