The script model_training_forecasting.py is applied for a combination of each algorithm, split date, and set of covariates. The split date defines the date on which the time series is split into training and testing data. The following example trains a TFT model on all data available until July 17, 2019 (inclusive), using the laboratory component (LC) covariates.
python model_training_forecasting.py 2019-07-17 LC TFT
Two example outputs for such an execution are provided in the folder out_TFT.
The script create_walk-forward_executions_NeSI.py has been implemented to create all combinations and creates lists of command line calls that are used to submit job arrays on the NeSI high-performance computing cluster. See the files in NeSI_execs for all the executions.
The forecasts resulting from the one-day walk-forward validation are first concatenated by the script evaluation/collect_and_concatenate.py. During the concatenation, the weekly trend forecast is created by averaging the daily forecasts. Concatenated files are written into the folder concatenated, providing all forecasts summarized by algorithm and covariate set.
The script compute_scores.py is then used to compute the evaluation metric (MAPE) for the machine learning benchmark.
The folder pearson_correlation_analysis provides the Python script needed to run the Pearson correlation analysis implemented for the laboratory component. This script also creates the heatmap shown in the manuscript.