Hierarchical-Time-Series-Modelling-MinT-

Hierarchical time series modelling is a concept that entails predicting coherent forecasts for time series data that have a hierarchical structure. Maintaining coherency across the hierarchy is a major challenge for hierarchical time series modelling. The Hierarchical Mixture Networks (HINT) framework is one framework that has been proposed for efficient and accurate coherent forecasting. This project aims to compare the performance of different reconciliation techniques in forecasting retail sales volumes using the HINT framework. The techniques chosen are the bottom-up and minimum trace algorithms which will be evaluated using the scaled continuous ranked probability score (sCRPS), root mean squared error (RMSE) and the mean absolute error (MAE) metric. From the results, we see that with the sCRPS metric, the bottom-up algorithm performed better than the minimum trace algorithm on a store basis, but the minimum trace algorithm performed significantly better on a total level compared to the bottom-up approach. According to both the RMSE and MAE metrics, the minimum trace algorithm performed better than the bottom-up technique on a store level and the bottom-up approach did slightly better at an aggregated level. It is difficult to determine which reconciliation algorithm performs better overall and but believe that the work done here could aid in future work to create such a methodology. This code is for the minimum trace algorithm and can be run on Google Colab. The 'neuralforecast' package will need to be installed using the following command: !pip install neuralforecast