Results of the pre-registered replication of the LTSF-Linear versus Transformer comparison with a matched tuning budget (arXiv 2205.13504): ETTh1 and ETTh2

Results (4 Nov 2026) of the private pre-registered replication whose frozen protocol is deposited at doi:10.5281/zenodo.23907285. Arm R (the authors' scripts compared with the printed Transformer numbers): Replicated under protocol X with a Repeat-baseline tolerance of 1.5 percent (a post-freeze deviation, disclosed in the PREREG; Inconclusive under the frozen 1 percent), median improvement 8.05 percent [7.53, 9.26]. Arm P (six models, K = 8 tuning configurations each, look-back included, selection by validation): Inconclusive under protocol X, median improvement 7.60 percent [6.04, 9.63] against a bound of 6.8 percent. On ETTh2 the linear model beats the best Transformer by 14.7-20.6 percent at every horizon; on ETTh1 the two are within about 1 percent at horizons 96 and 192 and FEDformer is better at 336 and 720 (by 2.8 and 4.0 percent). Contains the amended decision code (decide_amended.py; the frozen version is in the earlier deposit), the harness, the result, the selected configurations, all 424 job records, the forecast written before any result, and its scoring. Two datasets only; not tested: ETTm, Weather, Exchange-Rate, ILI, Electricity, Traffic, newer Transformers.