Reservoir Computing for Macroeconomic Modelling

With the increased globalisation of financial markets, financial crises and shocks have become more frequent and contagious. Despite a mass of data on financial markets and the banking sector being available, traditional macroeconomic models account for financial influences primarily through monetary aggregates and interest rates. This narrow representation of financial markets is proving increasingly inadequate, reflected in poor forecasting performance of traditional macroeconomic models.

A major weakness in the existing models is the absence of aggregate variables which can adequately represent broad financial market conditions. This absence has stimulated a growing literature on the construction of aggregate financial conditions indices (FCIs) formed from a large set of financial variables by use of dynamic factor models (DFMs). Prior research by the PI and Co-I have highlighted some of the limitations of DFM-based FCIs. Such FCIs do not perform well if the number of financial variables considered becomes large and have difficulties taking country specific characteristics into account. Due to these shortcomings, their predictive performance has been disappointing.

The project suggests applying new techniques in machine learning, namely reservoir computing, in the context of macroeconomic modelling and forecasting. Machine learning techniques have shown remarkable forecasting performance in dealing with large datasets, thereby potentially addressing the shortcomings of conventional factor-based FCIs. The project will draw on the expertise of an interdisciplinary and cross-sector research team of economists, statisticians, mathematicians, data and computer scientists and economic practitioners. The PI (a statistician) and Co-I (an economist) will team up with two leading experts in reservoir computing as research partners and the Ministry of finance, Greece as policy partner.

Reservoir computing is a relatively recent approach to machine learning and has not been applied to macroeconomic modelling and forecasting yet. The technique is particularly promising in the given context as it can deal with a mix of time series data and unstructured data, handle frequency mismatch and high dimensionality of input data well, and provides interpretable results, a key condition from an economic modelling perspective. Multiple empirical studies confirm exceptional pertinence of reservoir computing in various forecasting exercises with similar data types as in the macro-finance domain.

The identification of machine learning techniques that can improve macroeconomic modelling and forecasting by use of high dimensional input data that adequately represent financial market dynamics is potentially groundbreaking for economists and economic practitioners. To ensure maximum policy relevance of research outputs, the research team will involve academic beneficiaries and research users in the design, execution and dissemination of the research. All research outputs, including scoping studies, research briefings and journal publications as well as code written in Python, will be made available open source for researchers and research users to access.

The research will be conducted over 12 months in three phases. The first three months are dedicated to involving a wider research community and setting up a strategic network of economists, data and computer scientists, statisticians, mathematicians and economic practitioners. The subsequent six month of the project are dedicated to developing the application of reservoir computing to macroeconomic modelling and forecasting by the project team. The final three months of the project are dedicated to testing and competitively evaluating the performance of new methods in machine learning (including but not limited to reservoir computing) and other fields when applied to macroeconomic modelling and forecasting through the organisation of a 'hackathon' challenge.