Phongsakon/the-traders-trinity-checkpoints

The Trader's Trinity — Production Checkpoints

R6 (latest expanding-window round) production checkpoints for the Nordic day-ahead electricity price forecasting benchmark from The Trader's Trinity.

Trained 2026-04-08 with the residual-asinh target transform on hourly DK1/DK2 data spanning 2016-01-01 → 2024-06-30 (R6 train), validated on 2024-H2, tested on 2025-H1. See the GitHub repo's CLAUDE.md and research/AUDIT_LOG.md for the complete methodology.

Contents

Path	Model	Zone	Test MAE (R6)
production_dk1/lightgbm.pkl	LightGBM	DK1	27.99
production_dk1/xgboost.pkl	XGBoost	DK1	28.04
production_dk1/extra_trees.pkl	ExtraTrees	DK1	28.32
production_dk1/random_forest.pkl	RandomForest	DK1	28.59
production_dk2/lightgbm.pkl	LightGBM	DK2	29.24
production_dk2/xgboost.pkl	XGBoost	DK2	29.38
production_dk2/random_forest.pkl	RandomForest	DK2	29.34
production_dk2/extra_trees.pkl	ExtraTrees	DK2	29.36

Each .pkl is a Python pickle of {"model": <fitted estimator>, "meta": {...}} with metadata for the asinh scale, feature column list, hyperparameters, and the train/val/test split dates.

Usage

The simplest path is via the inference layer in the GitHub repo:

from research.ml.scripts.inference import ProductionForecaster

# Auto-downloads the requested checkpoint from this HF repo if not local
fc = ProductionForecaster(zone="DK1", model="lightgbm")
forecast = fc.predict_range(start="2025-07-01 00:00", end="2025-07-07 23:00")

Or pull a specific file directly:

from huggingface_hub import hf_hub_download
import pickle

ckpt_path = hf_hub_download(
    repo_id="Phongsakon/the-traders-trinity-checkpoints",
    filename="production_dk1/lightgbm.pkl",
)
with open(ckpt_path, "rb") as f:
    payload = pickle.load(f)
model = payload["model"]      # the fitted LightGBMForecaster wrapper
meta = payload["meta"]        # asinh_scale, feature_columns, train_end, ...

Citation / license

These checkpoints are released alongside the bachelor's thesis and journal manuscript "A Regime-Robust Benchmark of Forecasting and Reinforcement Learning Trading for Day-Ahead Electricity Markets: Lessons from the 2022 Nordic Energy Crisis" (Konrad, Adam, Ayvaz; in preparation, 2026). License: MIT.