Modelling time-varying interactions in complex systems: the Score Driven Kinetic Ising Model

We introduce a generalization of the Kinetic Ising Model using the score-driven approach, which allows the efficient estimation and filtering of time-varying parameters from time series data. We show that this approach allows to overcome systematic errors in the parameter estimation, and is useful to study complex systems of interacting variables where the strength of the interactions is not constant in time: in particular we propose to quantify the amount of noise in the data and the reliability of forecasts, as well as to discriminate between periods of higher or lower endogeneity in the observed dynamics, namely when interactions are more or less relevant in determining the realization of the observations. We apply our methodology to three different financial settings to showcase some realistic applications, focusing on forecasting high-frequency volatility of stocks, measuring its endogenous component during extreme events in the market, and analysing the strategic behaviour of traders around news releases. We find interesting results on financial systems and, given the widespread use of Ising models in multiple fields, we believe our approach can be efficiently adapted to a variety of settings, ranging from neuroscience to social sciences and machine learning.