Machine learning applications on time series data for systematic investing

  • Elizabeth Fons

Student thesis: Phd


This thesis studies the use of machine learning algorithms on financial time series data. Neural networks in particular are one of the most active areas of research in machine learning, and their application on financial investment is becoming more relevant each year. Still, several challenges remain, and one of the central problems of using neural networks for asset allocation is developing robust methods that generalise well and will work on real-world, unseen data. In the first part of the thesis we focus on enhancing an emerging trend in passive investment called smart beta. Whilst Smart beta strategies perform well in the long run, these strategies often suffer from severe short-term drawdown (peak-to-trough decline) with fluctuating performance across cycles. To address this, we build a dynamic asset allocation system using Hidden Markov Models (HMMs) and test it in a variety of portfolio construction techniques. The resulting portfolios show an improvement in risk-adjusted returns, especially on more return-oriented portfolios (up to 50% of risk-adjusted excess return). In addition, we propose a novel smart beta allocation system based on the Feature Saliency HMM (FSHMM) algorithm that performs feature selection simultaneously with the training of the HMM, to improve regime identification. We evaluate our systematic trading system with real life assets using investable indices; the results show model performance improvement with respect to portfolios built using full feature HMMs. We then focus on enhancing active investment strategies by proposing a machine learning framework for stock trading. Stock classification is a challenging task due to high levels of noise and volatility of stocks returns. In this work we present a series of methodologies to improve model generalization when using neural networks. We first introduce the use of data augmentation methods whose use in time series classification is still at an early stage. This is even more so in the field of financial prediction, where data tends to be small, noisy and non-stationary. We evaluate several augmentation methods applied to stocks datasets using two neural network models and the results show that several augmentation methods significantly improve financial performance when used in combination with a trading strategy. Furthermore, we introduce transfer learning and ensemble learning. Transfer learning can help improve stock classification, by pre-training a model to extract universal features on the full universe of the S&P500 index and then transferring it to another model to directly learn a trading rule. Transferred models present more than double the risk-adjusted returns than their counterparts trained from zero. In addition, we propose the use of data augmentation on the feature space defined as the output of a pre-trained model (\ie augmenting the aggregated time-series representation) and compare this augmentation approach with the standard one, i.e. augmenting the time-series in the input space. Finally, we move towards automating data augmentation, where one of the biggest challenges is how to search over the space of transformations, which can be prohibited given the large number of possible transformations and their associated parameters. We propose two adaptive, sample-specific methods that automatically select augmentation methods or weight their importance during training. We test this methods on time series datasets and on financial datasets building a trading rule.
Date of Award31 Dec 2022
Original languageEnglish
Awarding Institution
  • The University of Manchester
SupervisorJohn Keane (Supervisor) & Xiaojun Zeng (Supervisor)


  • Neural Networks
  • Hidden Markov Models
  • Time Series
  • Machine Learning
  • Systematic Investment

Cite this