Abstract/Summary

This thesis studies applications of deep learning in asset pricing. The thesis contributes to the prediction of equity returns, portfolio weights, and realised volatility, as well as to the interpretability of models through the measurement of variable importance, and the optimisation of hyperparameters using deep learning. This thesis presents three contributions. The first contribution is that we analyse forecast combination methods in the context of machine learning to predict equity returns. Whilst individual models lack robustness, forecast combinations over two levels display stability and have Sharpe ratios of up to 3.06 on data from 1987 to 2020. We use decision trees in genetic algorithms to analyse the structure of variable influence. The impact of these variables displays non-constancies and shows variations across different models and data. We propose a new performance measure for risk premium forecasts which leads to more robust evaluations than existing performance measures such as R2, whilst providing economic interpretability. This measure can be linked to the advantages models offer for portfolio choice. The second contribution we make is that we propose a new way of highlighting the context of information from past and present financial data to price equities in a universal approximation setup. The combination of adding a time wise lookback space and using contextual deep learning architectures with convolutional-transformer architecture enhances pricing in several ways. Using US stocks between 2003 and 2020, we improve the accuracy and performance of the models, increasing Sharpe ratios of managed portfolios from 0.91 in a simple feed forward benchmark to over 1.19 in our approach. The methodology gives insights into the factors that underpin pricing in the time and predictor space, making mid-term momentum factors less and short term momentum factors as well as financial statement scores more important than in the traditional approach. The third contribution we make is that we propose a deep learning framework for asset pricing that leverages multidimensional, mixed-frequency data with different durations, accounting for time-lags in non-lead frequencies. Our analysis forecasts weekly returns, realised volatilities, and optimal portfolio weights for the 500 largest US stocks using 143 financial indicators at weekly, monthly, and quarterly intervals. While our multi-kernel architecture offers flexibility in handling irregularly timed data, it did not surpass simpler models that use a unified frequency, which achieved a strong Sharpe ratio of 1.36. Despite efforts in hyperparameter optimization, the model’s complexity limited its performance, highlighting the potential for further enhancement in modelling mixed-frequency data.