Estimate the limit of predictability in short-term traffic forecasting: An entropy-based approach

Accurate short-term traffic forecasting is the cornerstone for Intelligent Transportation Systems. In past several decades, many models have been proposed to continuously improve predictive accuracy. A key but unsolved question whether there a theoretical bound accuracy with which can be predicted and that limit directly estimated from data. To answer this question, we use core concepts in information theory derive of predictability forecasting. Theoretical analysis proves conditional differential entropy poses rigorous lower negative-log-likelihood (NLL) probabilistic models. And continuous form Fano’s theorem further gives loose mean-square-error (MSE) deterministic Based on special properties dynamics, two assumptions are made estimate metrics: cyclostationarity (traffic phenomena show strong periodicity) localized spatial correlation (due kinematic wave propagation). They allow formulating as function longitudinal space time-of-day finds most uncertain locations periods solely Experiments univariate accumulation network-level speed selected models, including some state-of-the-art deep learning indeed cannot outperform bounds just approach them. The depends time-of-day, network locations, observation range, prediction horizon. results reveal stochastic nature dynamics improper prior distribution output major factors restrict performance. summary, method estimates trustworthy performance boundary These conclusions helpful studies domain.