Explainable Forecasts of Disruptive Events using Recurrent Neural Networks

Abstract

This paper describes the Crystal Cube method we developed for forecasting disruptive events around the world, specifically Irregular Leadership Change. Crystal Cube uses a Recurrent Neural Network (RNN) with Long-Short Term Memory (LSTM) units for forecasting. In this paper special emphasis is put on explanations of the network forecasts. We are using SHapley Additive exPlanations (SHAP) for individual forecast explanations and we are aggregating the explanations separately for True Positives, False Positives, True Negatives, and False Negatives. The method can be extended to Deep Reinforcement Learning models for self-driving cars or unmanned fighter jets.