LaCoCast: Traffic Flow Forecasting with Lag-Aware Cross-Sensor Corroboration

Traffic flow forecasting estimates future vehicle counts at monitored road locations from historical observations. A central challenge is determining how strongly forecasts should respond when recent observations depart from recurring traffic patterns. We propose LaCoCast, which combines a periodic reference with lag-aware cross-sensor corroboration to regulate local forecast corrections. The reference encodes temporal translation symmetry by assigning the same baseline flow level to matching phases across cycles, while a local forecaster predicts deviations from this reference. Earlier neighbor observations reconstruct recent target deviations, providing evidence for evaluating their agreement with observed target behavior. A correction gate combines this agreement with evidence availability, neighbor coverage, and local context to adjust the residual candidate’s contribution at each forecast horizon. Experiments on PEMS03, PEMS04, PEMS07, and PEMS08 yielded mean absolute errors 0.38–6.13% below the lowest published baseline values included in the comparison for each dataset. Ablation and evidence perturbation experiments further support the use of temporally matched neighbor observations to regulate residual contributions, connecting a periodic structural reference with adaptive traffic flow forecasting.