Machine learning in fixed income markets: forecasting and portfolio management

The fixed income market (i.e. bonds) is a massive asset class with an overall size of USD 100 trillion that remains relatively under-investigated using machine learning. The yield curve is its centrepiece for investors, regulators and the overall economy. In this thesis we apply machine learning to both bond forecasting and portfolio management. More specifically, we consider three different topics. The first two topics focus on machine learning models for forecasting, using multilayer perceptrons (MLPs) and long short-term memory (LSTM) networks, respectively. The third and final topic is on using reinforcement learning (RL) for portfolio management. These topics address specific gaps in the literature. In particular, existing literature lacks direct solutions for modelling the yield curve as a whole using machine learning. In addition, there is lack of work analysing and drawing interpretations from internal signals of black-box type of models like MLPs and LSTMs. Finally, there is no work that establishes RL as a framework for bond portfolio management. In more detail, for the first topic, two models are used for forecasting the European yield curve: multivariate linear regression and MLP, at five forecasting horizons. Five variants of MLPs are analysed, using different sets of features and, in some cases, including artificially-generated data from the linear regression model. We introduce a methodology relying on a rigorous feature selection process to identify the most relevant features, which we found to be different for each target and each forecasting horizon studied, reinforcing the importance of custom-built models. Considering all forecasting horizons, the results show that the MLP using the most relevant features achieve the best results and the addition of artificially-generated data tends to improve accuracy. Overall, the results demonstrate the superiority of MLPs to forecast the yield curve, when compared to benchmarks and other studies in the literature. For the second topic, we conduct a study of 10-year bond yield forecasting using dynamic LSTMs. In doing so, we study multiple temporal horizons, and compare the results to static MLPs using different covariates. Results show that the LSTM is capable of achieving lower prediction errors with higher confidence. Using a novel method we refer to as LSTM-LagLasso, we then go on to study the internal gating signal of a trained LSTM, and explain their dynamics using exogenous variables that can potentially influence bond price formation. By considering these variables at various lags and the Lasso method to select features, we show how different hidden units of the LSTM dynamically switch to make predictions during different temporal regimes and how their evolution is influenced by different external variables. Finally, for the third topic we develop an autonomous RL system which includes a purpose-built environment for fixed income portfolio management. It interacts with an agent with a state-of-the-art algorithm, the deep deterministic policy gradient (DDPG). We successfully test this system with environment inputs from four bond exchange traded funds (ETFs), using a novel methodology which involves a number of modifications in relation to the literature. The RL algorithm showed some signs of instability requiring further research. Despite this, we demonstrate how to extract the best agents from the system, during training, using principled selection criteria. With this selection, we are able to find agents capable of outperforming both the static buy and hold strategy and the best asset in the portfolio. Overall the results confirm the potential of RL for this application. In conclusion, this research covers three topics using machine learning in fixed income markets. In doing so, we have extended the state-of-the-art literature. The main overall objective is to provide financial practitioners with additional machine learning tools