Project structure:
data: Input data and intermediate files for modelingout: Prediction filessrc: Source codeutils: Utility functions used by training scriptsconfig: Configurations used by training scriptsrun_pipeline.py: Training script which takes data and a model config as input and writes predictions to theoutdirectorycompute_final_predictions.py: Imports one or more prediction dataframes, combines the outputs, performs final post-processing adjustments, and writes final predictions.
Run the run_pipeline script to generate predictions using a model config.
Additional model configs can be created and imported into this script to run
different models and save the prediction files with different names in the out folder.
Run the compute_final_predictions script to perform post-processing such as
combining predictions from multiple models and adjusting predictions to have
the same mean as the test set. This script outputs a "final" prediction dataframe.
- Random Forest is more accurate with most weather variables excluded - PCA and feature selection routines were not helpful - Manually adjusting the mean prediction to match the test set was helpful
- XGBoost significantly outperforms Random Forest and has several additional advantages: - Imputes missing values automatically - Results are close to the average of the test set without need for manual adjustment