The project focused mainly on studying datasets consisting of prices or values of financial instruments over time. Futures markets have been one key source of financial data, as these instruments are exchange-listed products in many cases offering high liquidity, which is a precondition for the implementability of risk management techniques in practice. Futures provide a unified framework that allows extracting the connectivity and the relations between the entities in a portfolio covering commodities, fixed income, equities and forex and, in general, have the ability of mapping a vast part of the world’s economic activity. Networks can be constructed by looking at correlations among price changes of different futures or assets. In this project, we investigated the dynamics of the networks. We tested investment strategies (cross-asset risk premia) inspired by or that are able to exploit the network models.
In recent years, research in data science and machine learning introduced Machine Learning (ML) Explanations: models constructed to explain learning algorithms outcomes. The explanations can help, for example, to fulfil fairness requirements when dealing with human/machine interactions or decision making.
To assess the quality of investment strategies, which is a fundamental part of the research in asset management, we introduced a new way to study their performances. The standard method of testing an investment strategy is by evaluating its performance retrospectively by using historical data (backtesting). Using synthetic or bootstrap data, one can enrich this analysis and simulate the strategy behaviour in different scenarios and data configurations, obtaining a statistical distribution of strategy performances. By leveraging the power of ML explanations, we constructed a pipeline that can test the strategy in multiple scenarios and connect back the performance to features of the historical data that produced it. The ML explanations provide insights to understand the behaviour of complex strategies: what are the actually relevant features and how do they impact the performance. Performance can be e.g. returns, volatility or a risk-adjusted measure whose behaviour can be very non-linear or path-dependent.
Our pipeline and results obtained by using it to investigate two prominent risk parity strategies – “equal risk contribution” and the network-inspired “hierarchical risk parity” – have been accepted for publication in the Journal of Financial Data Science.
In the project, we also introduced novel network-based diversification strategies that are able to go beyond linearity and long-only positions.
Moreover, two other papers have been published concerning the use of ML explanations in finance, in these cases in credit risk, and how networks can provide non-local and global insights into the ML model.
The project content and results have been disseminated also in workshops, conferences and events reaching researchers, investors, the general public and, thanks also to the synergy with the h2020 FIN-TECH project, financial supervisors and policy-makers.
