Traffic Forecasting Without Extra Training

A new study shows that large pre‑trained models can predict how many cars will be on a road without needing extra data or heavy training. Traditional deep learning tools get better as they learn from millions of traffic records, but that process is slow and costly. The researchers tested two popular pre‑trained models, Lag‑Llama and Chronos, on traffic data from a city. Both models were able to guess future vehicle counts accurately even when the patterns changed suddenly, such as during a holiday or an accident.The models that remember longer time windows and are bigger in size were more precise, though they took a bit longer to make each prediction. Importantly, these models can be dropped into new traffic situations right away because they were trained on a wide variety of data. The work suggests that using such foundation models could make traffic management faster and cheaper than relying on custom deep learning setups.