Causally interpretable multi-step time series forecasting: A new machine learning approach using simulated differential equations

This work represents a new approach which generates then analyzes a highly non linear complex system of differential equations to do interpretable time series forecasting at a high level of accuracy. This approach provides insight and understanding into the mechanisms responsible for generating past and future behavior. Core to this method is the construction of a highly non linear complex system of differential equations that is then analyzed to determine the origins of behavior. This paper demonstrates the technique on Mass and Senge's two state Inventory Workforce model (1975) and then explores its application to the real world problem of organogenesis in mice. The organogenesis application consists of a fourteen state system where the generated set of equations reproduces observed behavior with a high level of accuracy (0.880 r^2) and when analyzed produces an interpretable and causally plausible explanation for the observed behavior.