Replication package: Horizon, Deployment Objective and Evaluation Protocol Decide the Ranking -- a controlled benchmark of load forecasters on 300 buildings

Abstract

Replication package for Horizon, Deployment Objective and Evaluation Protocol Decide the Ranking: Evidence from a Controlled Benchmark of Load Forecasters on 300 Buildings. Eight forecasters -- DLinear, an LSTM encoder-decoder, PatchTST, iTransformer, XGBoost and three naive rules -- are compared on 300 ASHRAE Great Energy Predictor III buildings under one preprocessing pipeline, a per-horizon hyperparameter search of identical budget, five seeds and index-aligned forecast origins. The advantage of attention over recurrence is largest at one hour and smallest at one day, the opposite of the long-range-dependency rationale that motivates these models; at the day-ahead lead the three sequence architectures are equivalent within a margin declared in advance. Deployment objective, site and evaluation protocol each move the ranking further than the architecture does. This deposit contains the pipeline, the summary results behind every table and figure, and the figures themselves. What is included code/ -- the full pipeline: preprocessing and sampling, the five learned forecasters and three naive rules, the horizon, heterogeneity, equivalence and Diebold-Mariano analyses, the economic evaluation, and the figure scripts. results/e1/ -- the main result tree: per-model metrics, the per-step metrics (stored as per-horizon gzips; see README.md for the one-line recombine), Diebold-Mariano tests, the equivalence tests, and the economic evaluation. results/e3probe/, results/e4-e7 -- the gate and profile ablations, heterogeneity, weather ablation, interpretability and leave-one-site-out. figures/, figures_from_station/ -- every figure in the manuscript and its supplement. run_*.sh -- the run drivers. Three carry pre-registered decision rules above the run they gate. tools/split_oversize_csv.py, experiments/EXPERIMENT_PROTOCOL.md, README.md, SHA256SUMS. What is not included, and why ASHRAE GEPIII load and weather data (~2.4 GB). Redistributable only through the competition's Kaggle page under its terms. code/01_preprocess.py retrieves it; a Kaggle account and API token are required. It is not deposited here and must not be. Model checkpoints and per-building prediction files. Several GB, regenerable from the code, and close enough to a derivative of the GEPIII load series that depositing them would be doing indirectly what the competition terms forbid directly. The manuscript. Deliberately out of scope. This record is a software and results deposit, not a preprint. Reproducing -- see README.md. Training needs a CUDA GPU; everything downstream of the saved predictions runs on CPU. code/00_audit.py and the smoke tests run without a GPU and without the full download.