Training on the Future: A Delay-Aware Audit of Test-Time Adaptation for Time-Series Forecasting
Fentazi, Ameur and Ksentini · arXiv 2610.12232
The outcome is the verdict on the one claim below, under the protocol we fixed in advance. It is not a verdict on the paper as a whole.
Claim tested
RLS-Bank cuts stream MSE by at least 25% versus TAFAS and PETSA.
What we ran and what came back
On the ETTm1 benchmark under the authors' protocol, with five model checkpoints we trained ourselves on new seeds, the authors' RLS-Bank cut online forecast error by 48.9% against TAFAS (interval 45.0 to 52.5) and by 51.8% against PETSA (interval 48.8 to 54.5), against a bound of 25%, and was lower in 5 of 5 seeds. The paper reports 53% and 56%. On the authors' own seed numbers RLS-Bank reproduced to within 0.03. In a second arm, reported separately with no verdict, we used the reference repository's default data split on three seeds: the reductions were about 9% and 24%. The size of the advantage therefore depends on the split and protocol. TAFAS and PETSA were run at their defaults, not tuned. One dataset, one horizon, one model family.
What this result supports
Test-time adaptation gain holds on ETTm1 at d=24.
What it does not support
Other datasets and horizons.
Record
Protocol, deposited before the run: 10.5281/zenodo.23377294
The authors were sent this result. They have not reviewed this listing.
How the protocol works is on the method page. If you are an author of this paper and something here is wrong, write to customers@aegismind.app and we will correct it on this page.