Probabilistic Conflict Forecasting with Automated Machine Learning

Abstract:

This contribution is at the grid level. The idea is to generate point predictions using automated machine learning, then do a grid-search over parameterizations of different distributions to find the one that performs best. Models were compared using the CRPS across test windows for (1) different Box-Cox transformations on the dependent variable, (2) different sets of predictor variables, and (3) different distributions (Poisson, Negative Binomial, and Tweedie). To make bolder predictions and to make use of different strengths across each setup, ensembles of probabilistic models were built and compared across the different Box-Cox transformations. Two different models were submitted. The dorazio_log model forecasts the log of the dependent variable and then back-transforms to the original scale. This model is more conservative but scores a lower CRPS. Other Box-Cox transformations produced higher forecasted values which scored better for some grid-months. The dorazio_ensemble model uses forecasts from five different Box-Cox transformations, selecting which to use based on a separate grid-month forecasting model that was trained to predict when to use which transformation. All other factors between these models are the same, including common input features and the Tweedie distribution.

Authors:

Vito D’Orazio