Machine Learning in Fixed Income Markets
Lead Research Organisation:
University of Southampton
Department Name: Electronics and Computer Science
Abstract
The yield curve is the centrepiece in bond markets, a massive asset class with an overall size of USD 100 trillion that remains relatively under-investigated using machine learning. The objective of this project is to extend existing machine learning models to the fixed income asset class, by developing innovative applications of machine learning models and techniques for the financial industry, leading to better forecasting systems. Overall, it includes assessing and adapting existing techniques to model and forecast the yield curve (i.e. interest rates as a function of time to maturity) and evaluate potential benefits of multitask learning (i.e. simultaneous modelling of several dependent variables) when compared to single task learning, resulting from transformation of the problem into independent single-target processes. The main societal value of this research is directly related to the importance of bonds in pension fund and insurance portfolio holdings. But several sectors of society and industry could benefit directly from this research, including central banks' policy making, the asset and risk management industry, and academia.
Organisations
| Description | Contributions The main innovative contribution of this research has been the extension of existing machine learning models to the fixed income asset class that has not been comprehensively and extensively covered using these techniques, as it is the case with other financial asset classes. This research brings together techniques used separately in diverse fields and applications, for different objectives, finding a new application in fixed income markets. Specifically, the contributions to current state of the art are as follows: 1. We conduct the first comprehensive study on yield curve forecasting using multilayer perceptron models (MLP) and multitask learning techniques, showing that MLPs can be used to forecast the yield curve. 2. We introduce a comprehensive methodology, combining several techniques, for the application to a selected number of machine learning models. This methodology includes the following characteristics. First, we use a rigorous feature selection process, instead of relying on a pre-selection of individual features. This is a vital part of the methodology given that the most relevant features depend on both target yield to predict and forecasting horizon. Second, we use a moving window of training data to incorporate the most recent information as it becomes available in financial markets, and the retraining of models at every time step for enhanced results. This methodology makes models more flexible to changing market conditions. Finally, our dynamic cross-validation technique presents a number of advantages and innovative aspects in comparison with some of the most popular and widely used techniques. First, it introduces diversity in the training dataset through the bootstrapped samples. Second, it ensures a fair comparison of models, by forecasting yields at the same dates. Third, it guarantees forward-looking forecasting (using only past data to predict future data, unknown to the model). Fourth, it avoids any type of overlapping between training data and data used for calculating the forecasting errors, by using testing data to determine the metrics. Fifth and last, it enables a direct calculation of out-of-sample forecasting errors. 3. We expand existing methodologies by including synthetic data generated by the linear regression model in our neural network models, and show that they outperforms models where this data is not included. This is relevant for the possibility of using hybrid models incorporating data generated by industry-established models. 4. We conduct an innovative application of a deep learning model (LSTM) to bonds. The results are compared to memory-free multilayer perceptrons (MLP). Our results validate the potential of LSTM networks for yield forecasting. This enables their use in intelligent systems for the asset management industry, in order to support the decision-making process associated with the activities of bond portfolio management and trading. Additionally, we identify the LSTM's memory advantage over standard feedforward neural networks, showing that the univariate LSTM model with additional memory is capable of achieving similar results as the multivariate MLP with additional information from markets and the economy. 5. We go beyond the application of LSTMs, by conducting an in-depth study of the model itself (opening the black box), to understand the representations learned by its internal states. Such explanations of what black-box models learn is a popular topic of interest for certification and litigation purposes. In more detail, we extract and analyse the signals in both states (hidden and cell) and at the gates inside the LSTM memory cell. This is the first contribution to demystify the notion of black box attached to LSTMs using a technique which is fundamentally different from the most relevant ones found in the literature. Other studies are applied to a different type of recurrent neural network (Giles et al., 2001) or perform an external analysis of the model (Fischer and Krauss, 2018). 6. Following the extraction of signals at those locations, we proceed to explain the information they contain with exogenous economic and market variables. For that purpose, we develop a new methodology here identified as LSTM-LagLasso, based on both Lasso (Tibshirani, 1996) and LagLasso (Mahler, 2009). This methodology is capable of identifying both relevant features and corresponding lags. |
| Exploitation Route | Potential direct applications and beneficiaries: 1. Central banks' policy making (social, regulatory and macro value) In order to fulfil their responsibilities on monetary, macroprudential and microprudential policies, the Bank of England, publishes a research agenda discussion paper that shows the areas of research deemed important for the Bank's mission of maintaining monetary and financial stability. The present project guidelines fits under theme 4 of the Bank of England's research agenda (BoE, 2015). 2. Asset and risk management (social and industry value) The result of this research will make a positive contribution towards a more generalised use of machine learning techniques in fixed income markets. There are several institutions, industries or sectors that will benefit directly from better forecasting tools in this field: pension funds and life insurance companies; the asset management industry in general; and national agencies managing public debt. 3. Academic research knowledge (social and academic value) Academia will benefit from further research in the fixed income area to apply these new techniques, increasing the amount of published research. Moreover, there are several other potential areas of research where this knowledge could be applied within theme 4 of Bank of England's Research Agenda. |
| Sectors | Financial Services, and Management Consultancy,Other |
| URL | https://doi.org/10.13140/RG.2.2.10212.53129;https://doi.org/10.2139/ssrn.3415219;https://doi.org/10.1016/j.eswa.2018.11.012;https://dx.doi.org/10.2139/ssrn.3144622 |
| Description | This research is still active and further contributions are expected. We are encouraged by the positive response and interest that both the preprint and the journal article have generated from academia and the fixed income industry. In particular, a well-known international bank (UBS) and other financial institutions contacted us, demonstrating interest in this research. UBS requested permission to publish a summary of our research findings in the following technical report: Jessop, D.; Jones, C.; Gerken, J.; Winter, P.; Antrobus, O.; Stoltz, P.; Wu, S. (2018) Academic Research Monitor: Machine learning everywhere. UBS Global Quantitative Research, 25 July 2018. |
| First Year Of Impact | 2018 |
| Sector | Financial Services, and Management Consultancy |