A comparison of multitask and single task learning with artificial neural networks for yield curve forecasting

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. This paper is the first comprehensive study using artificial neural networks in the context of yield curve forecasting. Specifically, two models were used for forecasting the European yield curve: multivariate linear regression and multilayer perceptron (MLP), at five forecasting horizons, from next day to 20 days ahead. Five variants of the MLP were analysed with different sets of features: target to predict (univariate); the most relevant features; all generated features; and the former two incorporating synthetic data generated by the linear regression model. Additionally, two different techniques of multitask learning were employed: simultaneous modelling and transformation into multiple single task learning. The results show that considering all forecasting horizons, the MLP using the most relevant features achieved the best results and the addition of synthetic data tends to improve accuracy. Furthermore, different targets and forecasting horizons resulted in different relevant features, reinforcing the importance of custom-built models. In the two multitask learning methodologies no clear differentiation could be demonstrated, and several explaining factors are identified. Overall, the outcome is very encouraging for the development of better forecasting systems for fixed income markets.

More information

Identifiers

ORCID for Manuel Nunes: orcid.org/0000-0002-7116-5502

ORCID for Enrico Gerding: orcid.org/0000-0001-7200-552X

ORCID for Frank McGroarty: orcid.org/0000-0003-2962-0927

ORCID for Mahesan Niranjan: orcid.org/0000-0001-7021-140X

Catalogue record

Export record

Altmetrics