How Good Can Linear Models Be for Time-Series Forecasting?

The buried finding here isn't that linear models are good, it's that optimal context windows are non-monotonic across forecast horizons, meaning there's no single preprocessing recipe that generalizes. That specificity is what makes this actionable and also what limits how far the result travels beyond controlled benchmarks.

This connects directly to the co-failure ceiling paper from the same day ('When Does Combining Language Models Help'), which made a structurally similar argument: scaling through model combination hits a hard ceiling defined by shared failure modes, not by individual model capacity. Both papers, arriving together, push toward the same uncomfortable conclusion for vendors selling complexity as a solution. The forecasting paper applies that logic to a different axis, model size rather than ensemble count, but the underlying critique is identical. Neither paper argues simple methods always win; both argue the burden of proof for expensive architectures is higher than current practice demands.

Watch whether any of the major time-series foundation model providers (Nixtla, Amazon Chronos, Google TimesFM) publish rebuttal benchmarks on multivariate or irregular-frequency datasets within the next two quarters. If they don't engage directly, that silence is informative about how seriously they take the preprocessing-first framing.