The Thor model: an automatic nonlinear additive model for time series

Abstract

When facing an unknown forecasting problem, accuracy on the predictions as well as useful information about the underlying physics of the process are mostly appreciated. In this paper the Thor model, a fully interpretable model with automatic identification, is presented. Based on additivity assumptions and piecewise linear regression, it allows the analyst to gain insight about the problem by examining the automatically selected model. Monte-Carlo simulations have been run to ensure that the model selection procedure behaves correctly under weakly dependent data. Moreover, comparison over other well-known methodologies has been done to evaluate its accuracy performance, both in simulated data and in the context of short-term natural gas demand forecasting. Empirical results show that the accuracy of the proposed model is competitive against more complex methods such a neural networks.When facing an unknown forecasting problem, accuracy on the predictions as well as useful information about the underlying physics of the process are mostly appreciated. In this paper the Thor model, a fully interpretable model with automatic identification, is presented. Based on additivity assumptions and piecewise linear regression, it allows the analyst to gain insight about the problem by examining the automatically selected model. Monte-Carlo simulations have been run to ensure that the model selection procedure behaves correctly under weakly dependent data. Moreover, comparison over other well-known methodologies has been done to evaluate its accuracy performance, both in simulated data and in the context of short-term natural gas demand forecasting. Empirical results show that the accuracy of the proposed model is competitive against more complex methods such a neural networks.