Tag
FM-LLM propose a frequency-enhanced mixture-of-experts framework that adapts frozen LLMs to time series forecasting without textual prompts, achieving state-of-the-art results by injecting spectral representations and separating periodic/non-periodic decoding.
This preprint compares XGBoost and LSTM for forecasting transmitted heat energy in District Heating Systems, finding that XGBoost consistently outperforms LSTM while offering lower computational cost and environmental impact.
REATS is a new approach that uses LLM reasoning to perform interpretable, sample-adaptive ensemble learning for time series forecasting. It combines textual and numerical features with chain-of-thought reasoning and a two-stage fine-tuning framework, outperforming competitive baselines on eight benchmarks.
PhysAttNet is a physics-informed attention framework that augments lightweight CNN forecasters with domain-guided regularization to improve accuracy and generalization in industrial and astrophysical time series forecasting.
KReF introduces a training-free retrieval framework for long-term time-series forecasting that constructs empirical predictive distributions from similar historical lookback-future pairs, achieving strong CRPS performance across multiple benchmarks.
This paper identifies an objective gap in Direct Forecasting for multivariate time series and proposes CvLoss, a plug-in structural regularizer that constrains forecast residuals on a cross-variable graph to improve consistency across synchronous and asynchronous interactions. Experiments show consistent improvements over competitive forecasting models.
CAMP introduces a cycle-aware multi-scale patch mixer for time series forecasting, achieving state-of-the-art results on multiple benchmarks through adaptive cycle learning and horizon-guided patch refinement.
CastFSR is a Fast–Slow–Reflect agentic reasoning framework that leverages LLMs for context-aware time series forecasting, combining fast lightweight forecasters, slow deliberative reasoning, and reflective evaluation to improve forecasting accuracy and consistency.
The paper presents NRFormer+, a spatio-temporal Transformer for forecasting nuclear radiation using atmospheric diffusion guidance, achieving state-of-the-art accuracy on two large-scale benchmarks from Japan.
This paper proposes DiffDiff, a diffusion framework for probabilistic time series forecasting that embeds predictability asymmetry into the diffusion trajectory, outperforming six diffusion baselines on seven benchmarks across four prediction horizons.
This paper introduces a novel continual learning framework for time series forecasting that uses attention-guided experience replay to enable models to adapt to new data distributions while avoiding catastrophic forgetting, evaluated on benchmarks and real-world piezometric data.
This paper proposes Task-Semantic Field Factorization (TSF), an LLM-guided framework that uses offline semantic construction from process documents to enhance time-series forecasting and soft sensing in industrial processes. TSF reduces MAE by 6.4% on average with nearly negligible added parameters and inference overhead.
This paper introduces NEST, a framework using a regime-oriented mixture-of-experts to handle dataset-level distribution shifts in time series forecasting, achieving state-of-the-art on various benchmarks.
This paper formalizes the 'Granularity Paradox' in time-series forecasting, showing that finer temporal disaggregation improves in-sample fit but degrades out-of-sample accuracy due to recursive error propagation, and benchmarks multiple models across granularities.
TopoBrick is a training-free framework for zero-shot building IoT forecasting that uses building knowledge graphs and an agentic topology sampler to select target-specific exogenous variables. It outperforms strong zero-shot foundation-model baselines and remains competitive with fully trained building-specific models.
Introduces Evolutionary Feature Engineering (EFE), a framework that uses LLM-based evolution to automatically discover preprocessing transformations for structured data, improving time-series forecasting and tabular prediction accuracy while preserving interpretability.
This paper introduces StateFlow, a recurrent forecasting framework that extends the Variability-Aware Recursive Neural Network (VARNN) to long-horizon multivariate time series forecasting by using a dual-state recurrent backbone and a chunk-based decoder, achieving competitive performance against strong baselines.
Introduces an evolutionary neural architecture search framework (EvoTS) for discovering task-adaptive Transformer-like models for multivariate time-series forecasting. The approach uses a modular genome representation and achieves competitive performance on ETT benchmark datasets.
Proposes Hierarchical Temporal Fusion (HTF), an extension of the Temporal Fusion Transformer that integrates a coherence-aware loss function to ensure forecasts are consistent across hierarchical levels, achieving improved accuracy and coherence on benchmark datasets.
Proposes an end-to-end agentic pipeline combining SSA-LSTM forecasting, LSTM VAE anomaly detection, and LLM-based reasoning with dynamic retrieval to generate prioritized maintenance recommendations for appliance-level energy anomalies, achieving a 90.4/100 score on a 16-scenario benchmark.