Paper Proposes AliO Method to Fix Forecasting Output Drift

Long-term time series forecasting models are used to predict weeks or months ahead from a current data sequence, in settings such as weather forecasting and electricity consumption planning. A paper posted to arXiv on October 8, 2026 by Kwangryeol Park, Jaeho Kim and Seulki Lee identifies a problem in how these models behave that standard accuracy scores do not catch: when the same future timestamp is predicted from two different, overlapping input windows, the model often returns two different answers. The paper describes this as low output alignment, and argues that the fluctuation it causes undermines a model’s reliability even when its overall error rate looks fine.

The authors’ proposed fix is a method called AliO, for Align Outputs. Rather than changing how a model learns to minimize distance from ground truth, AliO works to reduce the gap between predictions for the same timestamp that come from different lagged input sequences, addressing both the time domain and the frequency domain versions of the mismatch. Because no existing metric was built to measure this kind of consistency, the paper also introduces TAM, the Time Alignment Metric, which the authors describe as scoring alignment between a model’s own outputs rather than scoring distance from the true value, the way an error metric such as MSE does.

The paper reports that AliO raises the TAM score, with the authors also stating that forecasting performance improves rather than trades off against alignment. Accepted at NeurIPS 2025 as a 46 page submission, the work sits in the growing body of forecasting research that treats reliability as a property distinct from accuracy.

The numbers

• TAM improvement: up to 58.2%, the reported gain in the paper’s alignment metric when AliO is applied.
• Forecasting performance: up to 27.5%, the reported improvement the authors say comes alongside the alignment gain, rather than at its expense.

The practical upshot for a team running a forecasting system in production is that a model can pass every standard accuracy check and still disagree with itself from one day to the next about what a given future date looks like. That gap is invisible to MSE and similar scores, so a procurement checklist built only around accuracy benchmarks will miss it entirely. Until alignment metrics such as TAM become part of standard vendor reporting, buyers evaluating forecasting tools for planning or trading decisions have no simple way to ask for this property by name, let alone compare it across providers.