The paper introduces series‑wise cross‑validation (CV) for global time‑series forecasting models, contrasting it with the traditional time‑wise CV that partitions data along the time axis. It formalizes several series‑wise CV schemes—K‑fold, bootstrapping, and Monte Carlo—and evaluates them on 11 datasets with 10 neural network architectures across five metrics, including performance estimation, overoptimism, model selection accuracy, regret, and downstream forecasting performance. The study finds that series‑wise CV improves downstream forecasting accuracy by better guiding model selection for generalization across both time and series, while time‑wise CV slightly outperforms it for estimating out‑of‑sample error.
arXiv:2602. 02288v3 Announce Type: replace Abstract: Current time-series forecasting models are primarily based on transformer-style neural networks.
arXiv:2510. 04487v5 Announce Type: replace Abstract: While accuracy is a critical requirement for time series forecasting, an equally important desideratum is reasonable forecast volatility across forecast creation dates (FCDs).
arXiv:2602. 16224v2 Announce Type: replace Abstract: Time series data are prone to noise in various domains, and training samples may contain low-predictability patterns that deviate from the normal data distribution, leading to training instability or convergence to poor local minima.
arXiv:2603. 15506v2 Announce Type: replace-cross Abstract: We argue that the current practice of evaluating AI/ML time-series forecasting models, predominantly on benchmarks characterized by strong, persistent periodicities and seasonalities, obscures real progress by overlooking the performance of efficient classical methods.
Deep learning methods have achieved state-of-the-art in time series forecasting, yet their accuracy varies considerably across samples, as some instances remain inherently difficult to predict. Reject option mechanisms, which allow models to abstain from high-risk predictions, are well established in classification and regression but underexplored in forecasting.
arXiv:2310. 20545v3 Announce Type: replace Abstract: We present a multi-task optimization approach based on a deep learning architecture for time series forecasting.
arXiv:2603.01820v2 Announce Type: replace-cross Abstract: Forecasting financial asset returns is challenging due to low signal-to-noise ratios, non-stationarity and heavy-tailed distributions. We ben...
arXiv:2512. 22702v2 Announce Type: replace Abstract: Deep learning models have grown popular in time series applications.
arXiv:2606. 27282v1 Announce Type: new Abstract: Time-series forecasting research has been moving steadily toward larger architectures, from specialized transformers to general-purpose foundation models, on the assumption that capacity is what unlocks accuracy.
Aurora‑X is a billion‑parameter time‑series foundation model designed for extreme forecasting tasks. It employs a progressive curriculum that starts with channel‑independent pretraining, then adds cross‑variable dependencies, variable context and horizon lengths, and optional future covariates during mid‑training. A variable‑resolution post‑training stage allows adjustable temporal spans per token at inference, while a pattern‑guided mixture‑of‑experts expands capacity through sparse activation and expert specialization. An implicit quantile network head predicts arbitrary quantiles, enhancing probabilistic forecasting flexibility. Experiments on GIFT‑Eval, TIME, FEV‑Bench, TFB, and DAG‑Bench show state‑of‑the‑art performance against both pretrained TSFMs and task‑specific supervised models.
Time-series forecasting research has been moving steadily toward larger architectures, from specialized transformers to general-purpose foundation models, on the assumption that capacity is what unlocks accuracy. We take the opposite position: most of the gap can be closed at far lower cost by tuning preprocessing rather than scaling models.
The paper presents an empirical benchmark of nine modern deep‑learning models for time‑series forecasting of smart‑meter energy consumption, evaluated on two publicly available datasets. It examines how historical input length, prediction horizon, and model architecture affect accuracy, finding that longer historical context improves performance up to a saturation point and that accuracy declines with longer horizons. The study also compares computational complexity, showing that lightweight architectures achieve similar performance to heavier models, and notes that model choice has limited impact across most demographic and household subgroups.