A correction layer for traffic forecasting models keeps their confidence estimates honest as conditions shift underneath them.
Researchers introduced TEGER, a correction layer that sits on top of an existing traffic forecasting model and continuously updates how confident that model should be in its own predictions. Instead of retraining the underlying network, TEGER uses a fixed map of how sensors across a road network relate to each other, then applies a statistical correction (Gaussian conditioning plus a rolling volatility estimate) using only the most recently observed errors. The method never touches the forecasting model's weights, so it can attach to an already-trained system, including a frozen foundation model. Across four traffic datasets, it improved 60-minute forecast accuracy over a baseline that only tracked time, and applied to the time-series model Chronos, it cut a standard error metric, CRPS_sum, from 0.1798 to 0.1736 without any fine-tuning of Chronos itself.
That distinction matters because most forecasting tools calibrate their uncertainty once at training time and leave it fixed, even though real-world traffic drifts: rush-hour patterns change, sensors degrade, and incidents behave differently over time. A cheap, retraining-free way to keep uncertainty current is relevant well beyond traffic: it is the same problem anyone deploying a frozen foundation model on messy, shifting data eventually runs into.
The gains here are real but modest, a few hundredths on a benchmark metric, and this is a preprint awaiting peer review, so treat drift-proof forecasting as a promising result, not a settled one.