Code for paper "Probabilistic Forecast Reconciliation with Kullback-Leibler Divergence Regularization" accepted in ICDM 2023 Workshop "AI for Time Series Analysis".
In this paper, we conducted experiments on three datasets. This code is for the Infant dataset.
You can run run.sh to reproduce the results in this paper.
Specifically,
preprocess.py: transfer the data to training instances and generate a hierarchical structure.train.py: train the deepar-hier model with specified hyperparameters.search_params.py: use grid search to tune hyperparameters, especially lambda.evaluate.py: evaluate "deepar-hier" and "deeper" on test database_hier.py: run other existing probabilistic forecast reconciliation methods.compare.py: compare "deepar-hier" and other methods, output CRPS and MCB results.
DeepAR related code refers to https://github.com/husnejahan/DeepAR-pytorch