On quantifying the role of exogenous macro-economic information in machine learning for modelling financial data

Abstract

Data from the financial markets are a source of challenging inference problems. Machine learning tools are increasingly used for the analysis of financial data. They are observed to provide more accurate models than classical analytical models that depend on specific assumptions. In this work, we ask if the inclusion of external (exogenous) macro-economic information into a model fitting procedure may be useful to improve the quality of analysis and predictions of financial time series. This dissertation explores this case by addressing several problems in empirical finance which are tackled by using a range of machine learning methods with exogenous macro-economic data.

First, we study a non-parametric approach to mapping the price of traded option contracts to the value of the underlying asset and the time to maturity. We explore if additional information would be helpful in improving this mapping. We show that this is the case, and further we show that there is a relationship between volume traded and volatility of an asset that is not apparent in the raw data, but it is seen through their influence on the prices of options. Then, we consider the non-negative matrix factorization (NMF) method and extend it with eXogenous information to specify a new model (XNMF). We present a learning algorithm for it and illustrate its better performance than NMF using equity prices and underlying macroeconomic variables. We show how residual signals arising in time series analysis can be explained by a sparse regression taken over related macroeconomic variables (the Kalman LagLasso model) to help in financial analysis. A comparison between stock index values and Bitcoin using this model illustrates clear underlying differences between them.

Finally, we study a powerful representation learning framework popular in machine learning (VAE) and extend it with inductive exogenous variable. Thus, we created a probabilistic XNMF (PAE-XNMF) that is able to generate financial data, with lower reconstruction error than a probabilistic NMF; and Recurrent Neural Networks, specifically, the Long Short Term Memory model (LSTM). We show that LSTM captures time series dynamics. Then we combined LSTM with attention mechanism to gain more interpretability of the influence of macro-economic data on predicting financial time series.

More information

Identifiers

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

Catalogue record

Export record