AIBRIEF
On a single point delay forecasting benchmark, the paper finds 92.7% of high accuracy temporal convolutional network (TCN) forecasts name a delay they functionally ignore. A single history slice diagnostic can expose the gap.
A forecast is supposed to read the past to predict the future. A new arXiv preprint argues that, for some widely used time-series models, a forecast can hit near-perfect accuracy while functionally ignoring the slice of history it claims to use.
The paper separates three questions: can the true delay be recovered from the data, does the model report it, and does the forecast actually use it. The hard one is the third. The authors design a "matched masking" test that forces the model to use only the history it names, then watch whether accuracy survives. On a point-delay benchmark, 92.7% of high-accuracy TCN forecasts and 55.4% of N-HiTS forecasts lose that protection. The model names one delay, but the prediction routes around it.
The paper's one-hot control, which pins the model to a single delay and tests whether the forecast still holds, is a check any operator or auditor can demand from a vendor before trusting a black-box forecast in load planning, inventory, or demand. Accuracy alone is not evidence the model is reading the right signals.
The result is preprint, not peer-reviewed, and the empirical numbers come from a single matched-masking task. Whether the same pattern holds in N-BEATS, Transformers, or production-scale forecasting pipelines is the next open question. The diagnostic is the contribution that travels.