This work proposes Memory-Based Utility-Aligned Residual Evidence with Calibrated Activation for Selective Correction of Traffic-Flow Forecasts (MURECAST), which treats historical residuals as candidate interventions and ranks first in dataset–metric comparisons.
Abstract
Short-term traffic-flow forecasting predicts road-network states, yet spatio-temporal predictors can leave structured, context-dependent residuals. They are seldom reused at inference, while similarity-based transfer can introduce corrections that increase error. We propose Memory-Based Utility-Aligned Residual Evidence with Calibrated Activation for Selective Correction of Traffic-Flow Forecasts (MURECAST), which treats historical residuals as candidate interventions. After freezing a context-aware base forecaster, MURECAST builds a static out-of-sample residual memory from an independent period. At inference, same-node and time-valid constraints delimit records, a forecast-visible utility estimator re-ranks them, and utility-weighted top-K aggregation forms a multi-horizon proposal. A chronological calibration split provides an empirical one-sided lower score for applying the proposal or retaining the base forecast. Across PeMS03, PeMS04, PeMS07, and PeMS08, MURECAST ranked first in 11 of 12 reported dataset–metric comparisons, attaining the lowest mean absolute error (MAE) and root mean squared error (RMSE) on all four datasets and the lowest mean absolute percentage error (MAPE) on three; relative error reductions over the strongest published results were 1.69–9.76%. Non-beneficial corrections represented 17–26% of accepted proposals versus 41–48% of all valid proposals. These results show that MURECAST reuses observed errors while concentrating intervention on corrections with lower observed non-beneficial risk under the evaluated chronological protocol.
Large language models use auxiliary conditioning to incorporate additional information into time-series forecasts. However, auxiliary conditioning does not consistently improve accuracy and can increase forecasting error when its influence is poorly matched to the observed temporal context. We propose Forecast-Aware Re...
Yi-Yang Zhang, Yu-Liang Zheng, Bing-Ran Li et al.· 0 citations
Equity-relevant news evolves through temporally dependent corporate events, making historical information useful only when event continuity, information availability, and transition reliability are modeled. Existing LLM-based financial agents incorporate historical evidence, yet they provide limited support for preserv...
Findings support explicit historical retrieval as an effective and interpretable complement to the building load forecasting and improve the corresponding base forecasters in 90 of 96 dataset-level comparisons.
Shi-Qiao Zhou, Huanbo Lyu, H. Schöner et al.· 0 citations
This work proposes ECHO, a black-box forecast correction framework that uses AnalogE to assess retrieved historical analogues together with the current context and base forecast, learning transferable corrections through cross-domain pretraining and dynamically combines the two corrections according to their observed f...
Bo-Ya Zhang, Xilin Dai, Yi-Lin Wang et al.· 1 citation
Multi-step traffic flow forecasting requires integrating recent observations with historical context, yet determining how that context should shape feature representations and contribute to different prediction steps remains challenging. We propose a memory-conditioned spatio-temporal framework, termed ForePatch-DST, t...
Qian-Xin Xie, Bin Pan, Jin-Feng Xu et al.· Electronics· 0 citations
Contextual specialization can improve forecasting accuracy, but a correction selected on one historical interval may become unreliable under temporal distribution shift. To address this issue, we propose RAEGL, a Risk-Aware Evidence-Gated Learning framework for selective contextual forecasting. RAEGL retains a validate...
Yi-Fan Guo· arXiv.org· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.