delivery eta prediction

Order fulfillment time prediction in e-commerce logistics faces challenges in modeling uncertainty and insufficient coverage by traditional rule-based approaches. To address these, this paper proposes the first distributional forecasting framework that jointly ensures statistical validity and business sensitivity. It introduces Conformal Predictive Systems and the Cross Venn-Abers Predictor—novel to fulfillment time prediction—to deliver rigorous coverage probability guarantees. We further design cost-sensitive point prediction rules to prioritize late-delivery detection. The method integrates spatiotemporal fine-grained feature engineering with tree-based and neural network models, and employs Venn-Abers calibration for principled uncertainty quantification. Evaluated on large-scale industrial data, our approach achieves theoretical validity in distributional forecasts, improves point prediction accuracy by 14%, and boosts late-delivery identification rate by 75%, significantly outperforming existing rule-based systems.

To address the challenges of resource-constrained predictive maintenance and delayed fault response in rural railway infrastructure, this paper proposes and implements a low-cost, low-power wireless intelligent monitoring system. The system deploys heterogeneous sensors at critical train and track locations to form a lightweight IoT network; introduces a security- and latency-aware data transmission protocol; and adopts a distributed backend architecture enabling edge–cloud collaborative processing. Furthermore, it innovatively integrates lightweight machine learning models—including LSTM and Random Forest—for structural health assessment and early fault prediction. Experimental evaluation demonstrates a fault identification accuracy of 92.3% and a 37% reduction in operational maintenance costs. These results validate the feasibility and effectiveness of end-to-end hardware–software co-optimization for intelligent railway operations in underdeveloped regions.

This paper addresses the remaining time prediction problem for medium-sized orders in aviation logistics outbound warehouses. We propose and empirically evaluate four classes of predictive methods—deep learning models (e.g., LSTM), gradient-boosted trees (XGBoost, LightGBM), and two shallow architectures—using a novel, large-scale, publicly available event log comprising 169,000 real-world process traces. Experimental results show that deep models achieve the highest accuracy; however, shallow models attain comparable performance (within 3% MAE difference) while reducing computational overhead by one to two orders of magnitude. This confirms the practical viability of lightweight models for real-time industrial scheduling. Key contributions include: (1) the first large-scale, labeled event log specifically designed for aviation warehouse outbound processes; and (2) an empirically validated benchmark establishing a lightweight, efficient prediction paradigm for time estimation in logistics automation.

This study addresses the poor predictive performance for long-tail high-latency cases in business processes by revealing that the root cause lies in high uncertainty stemming from right-skewed latency distributions and heteroscedasticity, rather than mere data imbalance. Based on empirical analysis across 14 event logs, this work proposes a novel uncertainty-aware modeling paradigm that leverages the positive correlation between latency and uncertainty to optimize identification mechanisms. Experimental results demonstrate that this approach significantly improves the detection of critical high-latency cases, overcoming limitations inherent in traditional models. Consequently, this research offers a new perspective for predictive process monitoring that effectively balances accuracy with reliability, providing a robust solution for managing extreme latency scenarios in complex operational environments.

This study addresses the limitation of existing vessel estimated time of arrival (ETA) models, which are typically restricted to the next port of call and rely heavily on real-time AIS data, thereby struggling to forecast transit durations for future voyage segments. To overcome this, the authors formulate future-port ETA prediction as a segment-level time series forecasting problem and propose a Transformer-based multi-task learning framework that jointly predicts segment-wise sailing durations and port congestion states. A novel causal masked attention mechanism is introduced, coupled with shared latent representations, to mitigate the high uncertainty inherent in long-horizon predictions. The model integrates historical sailing durations, proxy indicators of port congestion, and static vessel attributes. Evaluated on a global real-world dataset from 2021, it significantly outperforms baseline methods, reducing MAE and MAPE by 4.85% and 4.95% compared to sequential models, and by 9.39% and 52.97% against gradient boosting machines, respectively.

Retail demand data are often plagued by strong seasonality, irregular spikes, and noise, which undermine the accuracy of traditional forecasting methods and hinder effective supply chain decision-making. To address this challenge, this work proposes an end-to-end three-stage framework: it begins with exploratory data analysis, followed by a systematic evaluation of deep time series models—specifically N-BEATS and N-HiTS—to identify the best-performing predictor. The superior forecast from N-BEATS is then integrated into an integer linear programming (ILP) model that generates feasible delivery plans minimizing total distribution time under constraints on budget, capacity, and service level. By combining high-accuracy deep learning forecasts with interpretable constrained optimization, the approach successfully translates four-week demand predictions for 1,918 units into cost-optimal, executable logistics plans, substantially enhancing operational efficiency.

This work addresses the challenge of evaluating model outputs in scenarios where ground-truth outcomes are delayed, censored, or private, rendering conventional code-based deterministic evaluation methods ineffective for immediate validation. The authors propose RouteCast, a novel framework that enables autonomous generation of auditable provisional prediction scores through typed, staged route modeling, reference-based analogy, and deterministic transformations, thereby supporting traceable and decomposable assessment of strategic pathways. Evaluated on 21 retrospective cases, RouteCast achieves an AUC of 0.756—significantly outperforming blind-evaluated large language models (AUC = 0.678) and performing comparably to identity-revealed LLMs (AUC = 0.761)—demonstrating its effectiveness and feasibility in settings with delayed ground truth.

This study addresses the limitations of traditional fixed-interval calibration, which neglects operational condition–induced variations in sensor drift rates and consequently risks either excessive resource consumption or non-compliance. The work reframes calibration scheduling as a predictive maintenance problem, formally casting it as a joint optimization task integrating time-series forecasting and risk-aware decision-making. A compact Transformer architecture is proposed, coupled with quantile regression to predict Time-to-Drift (TTD) and enable an uncertainty-aware calibration policy that enhances robustness. Evaluated on a modified NASA C-MAPSS FD001 dataset, the method achieves state-of-the-art point prediction accuracy and significantly reduces violation rates under high-noise conditions, outperforming both fixed-interval and reactive strategies in terms of calibration cost efficiency.

This study addresses the challenge of executing large orders in continuous double-auction markets under time and liquidity constraints. The authors propose a risk-constrained model predictive control (MPC) framework that dynamically optimizes trading decisions via quadratic programming while tracking benchmark schedules such as TWAP or VWAP. The approach permits strategic deviations from the benchmark to minimize expected execution cost, explicitly incorporating benchmark residual cost into the objective function to enable modular, data-driven deployment in live trading environments. Empirical evaluation using six months of NASDAQ Level 3 data demonstrates that the method reduces execution schedule gaps by 40–50% compared to cross-price benchmarks and significantly mitigates slippage. Performance improves further when integrated with price forecasts.

This work addresses the lack of established scaling laws and extensible foundation models in time series forecasting by proposing a unified training framework that yields a family of foundation models ranging from 4M to 2.5B parameters. Through a consistent architecture, large-scale data, standardized training protocols, and an innovative u-muP hyperparameter transfer method, the study provides the first systematic empirical validation that model performance in time series forecasting consistently improves with scale. The resulting models achieve state-of-the-art results across three major benchmarks—BOOM, GIFT-Eval, and TIME—and five checkpoints are released under the Apache 2.0 license to support further research and reproducibility.