Verified, then logged.
Sun 22 Nov 2026 · 00:24 KST / 15:24 UTC한국어 | ENSearch /

Breaking01 / AI

Longer histories improve forecasts of customer behaviour

The paper reported that extending the observation window improved forecast scores by more than the differences observed between sequence model families.

A language-model evaluation monitor compares two response timelines with warnings appearing before and after repetitive output begins. AI-generated image

Researchers reported on 20 Nov 2026 that forecasts of customer behaviour improved when models used 16 weeks of customer records.

The paper defines the observation window as the length of the history a model sees before the week it forecasts from. It evaluated performance using the coefficient of determination on a log scale, a measure of how much variation a model explains.

In the paper's 1M tier, extending the observation window from 8 to 16 weeks raised the score by 0.070 to 0.107 across all model families.

The researchers reported larger gains when they compared forecasts for the same customers in the same weeks. Scores also rose with a longer window on five public datasets.

With the observation window fixed at 16 weeks, the paper found that scores differed by at most 0.0052 across sequence model families.

The study did not establish how much of the gain came from the length of the window itself, independent of the calendar weeks it covered.