This repository contains implementations of hierarchical time series forecasting methods from the following papers:

• Baseline: Optimal forecast reconciliation for hierarchical and grouped time series through trace minimization
• Baseline: Regularized regression for hierarchical forecasting without unbiasedness condition
• Baseline: Simultaneously Reconciled Quantile Forecasting of Hierarchically Related Time Series
• Dynamic Combination of Heterogeneous Experts for Hierarchical Time Series (in submission to ICASSP'23)

To setup conda environment for running the program, use the following command:

conda env create -f htsf.yml

Activate the new environment:

conda activate htsf

This repository compares forecasting performance across benchmarked hierarchical time series (HTS) approaches on various real-world and simulated hierarchiclly related time series data.

HTS is a special type of multi-variate time series which has a predefined hierarchical structure between each variant. Below is an example of HTS with 13 variants, each variant could be e.g., a univariate time series that describes sales record at the corresponding level.

Input format on this example:

• HTS data: a Pandas DataFrame object, use numbers within bracket as feature names.
• Hierarchical structure: a list that includes the number of children of each non-leaf vertex: [[3], [4, 3, 2]].

• Australian Labour Force
• Predict Future Sales
• Web Traffic Time Series Forecasting
• M3 Competition Data
• AEdemand Data

• Bottom up (BU)
• Trace Minimization (MinT) including shrinkage, sampling and OLS estimators
• Empirical risk minimization (ERM)
• HTS prophet