BellmanPredict

Hurdle Model Combination

Binary classification probabilities and regression forecasts are blended using the conditional mean of a zero-truncated distribution. Given a classifier probability p and an unconditional mean forecast \mu_u, the zero probability is P0 = (\kappa / (\kappa + \mu_u))^\kappa. The conditional mean \mu_c is then \mu_u / (1 - P0) and the final prediction becomes::

y_hat = ((1 - \epsilon_{leaky}) * p + \epsilon_{leaky}) * \mu_c

During training the model jointly learns the mean \mu and shape \kappa of a truncated negative binomial distribution, while the binary head employs focal loss. The leakage term epsilon_leaky, configurable via PATCH_PARAMS, stabilizes the blend by keeping a small positive probability even when the classifier predicts zero.

Preprocessing

SampleWindowizer.build_lgbm_train now vectorizes target generation and unpivoting, producing the same dataset as the previous row-wise implementation but more efficiently.

Baseline Forecasting

Run baseline models with the provided configuration:

python LGHackerton/train_baseline.py --config configs/baseline.yaml --model naive

PatchTST Grid Search

Run a grid search over PatchTST settings:

python LGHackerton/tune.py --task patchtst_grid --config configs/patchtst.yaml

PatchTST Hyperparameter Selection

train.py determines PatchTST settings using the following order:

1. If artifacts/patchtst_search.csv exists (created by the grid-search task), the combination with the lowest val_wsmape is chosen. The associated input_len is applied to window generation and to the model.
2. Otherwise, Optuna results from artifacts/optuna/patchtst_best.json are used when available.
3. If neither artifact is present, default parameters from PATCH_PARAMS are used.

This avoids confusion when both grid-search and Optuna artifacts may exist.

Model and Tuner Registry

Training scripts resolve trainer classes and Optuna-based tuners through lightweight registries. Providing an unknown name raises a ValueError listing available options. See docs/registry_tuner.md for full usage and error-handling rules.

Combining Predictions

After generating predictions from different models, use the postprocessing utilities to ensemble them and create a submission file:

python LGHackerton/predict.py --model patchtst --out patch.csv
python LGHackerton/predict.py --model lgbm --out lgbm.csv

python - <<'PY'
import pandas as pd
from LGHackerton.postprocess import aggregate_predictions, convert_to_submission

preds = [pd.read_csv('patch.csv'), pd.read_csv('lgbm.csv')]
ens = aggregate_predictions(preds, weights=[0.7, 0.3])
convert_to_submission(ens).to_csv('submission.csv', index=False, encoding='utf-8-sig')
PY

The helper functions aggregate_predictions and convert_to_submission ensure consistent formatting and handle any missing or duplicate entries.

Documentation

• Callbacks
• Hyperparameter Tuning
• Model & Tuner Registries
• PatchTST Model