Machine Learning techniques for equity markets: from custom loss functions in predictive modelling, to dynamic neural network ensemble learning for capital deployment

Matuozzo, Alberto (2024) Machine Learning techniques for equity markets: from custom loss functions in predictive modelling, to dynamic neural network ensemble learning for capital deployment. PhD thesis, Birkbeck, University of London.

Abstract

This thesis proposes the adoption of machine learning methods for equity market forecasting, expanding the toolbox available to market participants. Machine learning techniques, while effective in extracting patterns, capturing non-linear relationships among different variables of different nature, cannot simply be ported to the financial domain, given its peculiarities. This research proposes solutions aiming to enhance domain knowledge rather than replace it. The stages of machine learning deployment are examined almost in parallel with the process that would span from formulating a forecast to deploying capital. The starting point of this study is a dataset composed by known financial time series covering over a decade. A heterogeneous set of features is developed applying Shannon’s entropy and transforming volatility indices in innovative ways to gauge trend and sentiment persistence. Effective equity market forecasting has to reconcile point estimation with direction. This study introduces a penalisation term in loss functions, in order to adapt for this purpose a regression framework. Given the sequential structure of data, and a multivariate setting, architectures based on convolutional neural networks over two dimensions are deployed, treating the obtained feature map similarly to an image. In this domain it is key to reconcile the importance of updating models with information pertaining the economic cycle. A novel ensemble technique based on a mixture of experts and a time-period classifier is developed for this purpose. This architecture is applied successfully to capital deployment problems. This research journey ends proposing a methodology to integrate machine learning in a human-led investment process: a domain expert selects a time period for training, considered appropriate given the economic outlook. Experiments have been performed on European markets, being developed, but under searched. The road less travelled is often more fruitful in search of alpha.

Metadata

Statistics

DownloadsShow export options

Export as XMLJSONCSV

Activity Overview

6 month trend

0Downloads

6 month trend

42Hits

Additional statistics are available via IRStats2.

Archive Staff Only (login required)

Edit/View Item