Essays on Machine Learning in Volatility Forecasting

Student thesis: Phd


This thesis attempts to take volatility research to a new area by exploiting advances in machine learning (ML), natural language processing (NLP), and Big Data analytics. The data underlying the three essays includes high-frequency limit order book (LOB) and daily business news archive for the 23 most liquid NASDAQ stocks over the sample period from 27 July 2007 to 27 January 2022. From 5-minute prices in LOB, I construct the daily realised volatility (RV) of these stocks, which becomes the target for RV forecasting in subsequent empirical studies. In the first essay, I compare the forecasting power of the HAR-family of models, the state-of-the-art go-to group of models for RV modelling and forecasting. I need the best-performing econometric model to serve as the benchmark to compare with ML and NLP models. In the process, this study also becomes the first to test the forecasting power and information content of major LOB variables and news sentiments compiled using the LM dictionary from Loughran and McDonald (2011). After a comprehensive analysis of different model specifications, I conclude that CHAR (continuous HAR) is the best-performing HAR model and that the forecasting patterns are distinctively different between normal volatility days and high volatility days. On high volatility days, a simple news count outperformed all LM news sentiment measures for forecasting RV, and LOB depth outperformed the LOB slope. On normal volatility days, the ask-side of the LOB is more price informative, suggesting that buy-side trades dominate under normal conditions. The second essay focuses on the forecasting power of ML, more specifically, long short-term memory (LSTM) as a pioneer of sequence modelling. 132 LOB variables, nine news variables and 6 HAR variables, together with a variety of model specifications, are tested in this essay. Results show ML outperforms all HAR models (including CHAR) on 90% of the out-of-sample period. Also, SHAP, an explainable AI technique, suggests mid prices, average bid and average ask at all LOB levels improved RV forecasting performance over time. The third essay develops a financial word embedding (FinText) from 15 years of business news archives and compares it to general word embeddings for RV forecasting using stock-related news and general hot news in a simple yet flexible NLP structure. Results show FinText produces more accurate RV forecasts than general word embeddings and significantly improved RV forecasts on high volatility days. This forecasting improvement carries over to normal volatility days when general hot news is used. Again, using SHAP, I was able to compile a list of key phrases from stock-related news and general hot news that moved stock volatility.
Date of Award1 Aug 2023
Original languageEnglish
Awarding Institution
  • The University of Manchester
SupervisorSer-Huang Poon (Supervisor) & Yoichi Otsubo (Supervisor)


  • Realised Volatility Forecasting
  • Machine Learning
  • Natural Language Processing
  • Big Data
  • Explainable AI

Cite this