Tag
GNBAN is a new graph-based neural architecture for long-horizon retail demand forecasting that combines heterogeneous graph learning with an interpretable basis-decomposition forecasting head, achieving 4-5% improvement on Walmart and Favorita benchmarks.
Proposes TempoWave, a plug-and-play temporal wavelet digit interface that maps time series observations into digit-wise embeddings from multi-wavelet coefficients, improving LLM-based time series forecasting and achieving state-of-the-art on multiple benchmarks.
PMDformer introduces patch-mean decoupling and specialized attention mechanisms to improve shape similarity modeling in long-term time series forecasting, outperforming existing methods on multiple benchmarks.
This paper introduces TS-Fault, a benchmark for evaluating time series forecasting models under structured fault scenarios like broken dependencies and regime changes, finding that clean-data accuracy often anti-correlates with robustness and that foundation models are especially fragile.
MorphStrata introduces a layer-specific stochastic noise injection strategy for generating diverse student models in a Moving Target Defense framework to enhance adversarial robustness in time-series forecasting, achieving up to 97.97% improvement in RMSE under BIM attacks with minimal training overhead.
Proposes SERAF, a multimodal retrieval-augmented framework for time series forecasting that uses both numerical similarity and self-generated textual descriptions to retrieve historical patterns, improving forecasting under non-stationarity. Experiments on seven real-world datasets show effectiveness over state-of-the-art baselines.
This paper evaluates LLM-based strategies (embedding, prompt, hybrid) against classical tabular models on an industrial car retrofit prediction dataset with hashed categorical features. It finds that tree ensembles outperform LLMs overall, but embeddings and hybrid approaches remain useful, while direct prompting fails without semantic cues.
Introduces SpikF-GO, a spiking neural network model for multivariate time series forecasting that combines graph-based inter-variable dependency modeling with spike-driven spectral processing, achieving state-of-the-art results among SNN methods with reduced energy consumption.
SPDM introduces a geometry-aware state-space model that uses manifold constraints on the symmetric positive definite manifold for time series forecasting, achieving state-of-the-art performance on 11 benchmarks.
Proposes EVIDENT, a framework that integrates Bayesian training and evidence-based ranking for neural architecture selection, demonstrated on subject-specific blood glucose forecasting in type 1 diabetes, systematically selecting low-capacity models that generalize reliably.
SARAF is a Stationarity-Aware Retrieval-Augmented Forecasting framework that adaptively balances relevance and diversity in retrieval for time series forecasting, modulating diversification strength based on dataset-level stationarity to handle non-stationary regime shifts. Accepted to KDD 2026, it demonstrates competitive performance over strong baselines on eight real-world datasets.
This paper theoretically and empirically examines adaptive patching for time-series Transformers, deriving conditions under which content-adaptive tokenization should outperform tuned uniform patching. Controlled experiments on standard benchmarks show that a well-tuned uniform baseline is competitive with dynamic patching methods, challenging the assumed benefit of adaptive approaches.
This paper introduces a framework for time series forecasting that uses importance-aware news compression and process reward model-guided retrieval to incorporate long news articles within fixed context limits, improving prediction accuracy across finance, energy, traffic, and Bitcoin benchmarks.
The paper proposes BatteryMFormer, a multi-level Transformer for early battery degradation trajectory forecasting that integrates aging-condition-aware decoding, meta degradation pattern memory, and dual-view encoding to capture multi-level degradation structures and SOC-localized variations, consistently outperforming state-of-the-art baselines across four battery domains.
This paper investigates whether high-quality Natural Language Explanations (NLEs) generated by LLMs from XAI outputs actually improve task performance, finding they do not aid accuracy but inflate confidence, revealing a quality-usefulness gap.
This paper presents an applied evaluation of foundation models for time series forecasting compared to supervised approaches across four operational domains, and proposes a Complexity Router to selectively assign series to the optimal model class for balancing accuracy and inference cost.
Proposes PESD-TSF, a physics-inspired structured decomposition framework for long-term time series forecasting that addresses periodic perception degradation, trend-noise entanglement, and loss of cross-variable dependencies via multiplicative periodic gating, multi-scale structured encoder, and cross-scale collaborative attention.
Google's new paper Nexus proposes transforming time series forecasting from statistical extrapolation to multi-agent reasoning, improving prediction accuracy via event context, achieving an 86.6% reduction in MAPE on the Zillow dataset.
Google's Nexus paper proposes an agentic framework that incorporates contextual events alongside numerical data for time series forecasting, achieving an 86.6% MAPE reduction on Zillow tests compared to direct chain-of-thought prompting.
DataDog releases Toto 2.0, an open-weights family of time series foundation models ranging from 4M to 2.5B parameters, demonstrating consistent scaling improvements and achieving state-of-the-art results on multiple benchmarks including BOOM, GIFT-Eval, and TIME.