Advancing Systematic and Factor Investing Strategies using Alternative Data and Machine Learning

Happersberger, David and Nolte, Ingmar and Lohre, Harald (2022) Advancing Systematic and Factor Investing Strategies using Alternative Data and Machine Learning. PhD thesis, Lancaster University.

Text (2021DavidHappersbergerPhd)
2021DavidHappersbergerPhd.pdf - Published Version
Download (24MB)

Abstract

This thesis advances systematic and factor investing strategies using alternative data and machine learning techniques. The first chapter studies the relevance of high-frequency news data for low-frequency factor investing strategies. We build various news-based equity factors for an investable global equity universe to investigate the factors’ ability to extend the information inherent in standard factor models. Specifically, we document that incorporating news-based equity factors benefits multi-factor equity investments, employing diversified multi-factor equity allocations but also more dynamic factor timing strategies. The second chapter examines dynamic asset allocation strategies that focus on explicit downside risk management. We investigate suitable risk models that best inform tail risk protection strategies. In addition to forecasting portfolio risk based on standalone models such as extreme value theory or copula-GARCH, we propose a novel expected shortfall (ES) and value-at-risk (VaR) forecast combination approach that utilizes a loss function that overcomes the lack of elicitability for ES. This forecast combination method dominates simple and sophisticated standalone models as well as a simple average combination approach in terms of statistical accuracy. While the associated dynamic risk targeting or portfolio insurance strategies provide effective downside protection, the latter strategies suffer less from inferior risk forecasts, given the defensive portfolio insurance mechanics. The third chapter extends the above ES and VaR forecast combination approach using machine learning techniques. Building on a rich predictor set of VaR and ES forecasts from an array of econometric models (including GARCH, CAViaR-EVT, dynamic GAS and realized range models), we leverage shrinkage and neural network models to form combination forecasts. Such machine-learned VaR and ES forecasts outperform a set of competing forecast combination approaches in terms of statistical accuracy as well as economical relevance in dynamic tail risk protection strategies.