Abstract
Synthetic data are often used to complete short building metering histories. We re-evaluate a transparent hour-of-week, temperature and AR(1) generator, calibrated on held-out buildings, completing 5–20% real training windows of 82 education buildings from the Building Data Genome Project 2 over 20 random splits. With the real window over a year before the test period, the supplement cut Prophet’s median paired error by 205.1 percentage points at 5%. With a matched forecast origin the reduction fell to 51.6, 15.1 and 3.1 points at 5, 10 and 20%, reversed over the first week, and did not differ detectably from white-noise, dispersion-matched or fully permuted complements (equivalence within ±3 points in only 4 of 18 comparisons). Real-only Prophet improved sharply when the same calendar weeks moved a year nearer the test period; season effects were imprecise. Flat-trend Prophet and a time-of-week profile on real data alone beat every supplemented model at five origins, suggesting the supplement mainly regularises Prophet’s trend. Supplements lowered Isolation Forest F1 through the contamination threshold, not average precision, and gave a profile-residual detector no supported advantage over regularising it without added data. Supplements should be tested with matched origins, structure-free controls, simple baselines and regularised models, per task and model.