wca-predictions

WCA Time Prediction

Machine learning system that predicts future World Cube Association competition results using an ensemble of time-series models. It is a part of this Github repository: https://github.com/Dhruvacube/wca-time-prediction

Trained Model Artifacts

This directory (training/models/) stores the trained machine learning model artifacts generated by the WCA Time Prediction training pipeline (train.py).

Directory Structure

Model artifacts are grouped into subdirectories named by their snapshot creation date (e.g., YYYY-MM-DD). The system automatically loads the latest available snapshot during inference unless a specific date is requested.

Inside each snapshot directory, you will typically find:

• lgbm/ Contains the serialized LightGBM model weights (.txt or .pkl) which have been trained on handcrafted tabular features.

• gru/ Contains the saved PyTorch model weights (e.g., .pt or .pth) for the Deep Sequence GRU model, trained on long-term chronological sequence data.

• ensemble_weights.json Stores the optimal model weights learned using constrained linear regression. These weights determine how the Classical, LightGBM, and GRU model predictions are combined to produce the final output.

(Note: The Classical forecasting models, like ARIMA/ETS, are often fitted dynamically at runtime on the provided history and may not generate large static artifacts here.)

Usage

You do not need to modify these files manually.

• To generate a new snapshot, run the main training script: python train.py.
• To load these models in your code, use the ModelStore class from models/model_store.py.