Economic networks and transaction data

Explaining the science

When analysing high-dimensional panels of macroeconomic or financial time series state-of-the-art methods tend to suffer from at least two main limitations.

Firstly, they often consider only linear dependencies. In contrast, real such time series are interconnected through a complex network of dependencies, and this network can display systemic risky behaviour. Secondly, the way in which factors affect data is seldom modelled, and typically only fully pervasive factors that affect all time series with similar magnitude are considered. In contrast, real such time series are collections of heterogeneous time series with different types of series influences by a diverse range of factors. 

Furthermore, traditional theoretical models in economics and finance assume that in large ecosystems the influence of a few individual agents is negligible. Yet in large economic and financial panels there often exist one or more subsets of time series that influence the entire cross-section. For example, with a few time series leading on other time series which can be viewed as lagging behind the leading series. One potential modelling approach is to suppose that the set of agents can be partitioned into clusters, such that within clusters there are no consistent lead-lag relationships, while across clusters there exist stable lead-lag relationships. 

The envisioned research will combine methods from machine learning and network analysis to reveal such hidden factors in the analysis of two types of high-dimensional data sets, with the aim of revealing complex relationships inherent within the data and improving forecasting accuracy. The first type of data is a network capturing pairwise interactions, such as cross-sectoral networks of input-output connections. The second type is a panel of economic time series, where the network structure can be examined by considering the correlations between different data streams.