📈 Time Series¶
🧠 NeurIPS2026 · 95 paper notes
📌 Same area in other venues: 🎞️ ECCV2026 (5) · 📷 CVPR2026 (7) · 🔬 ICLR2026 (121) · 💬 ACL2026 (8) · 🧪 ICML2026 (45) · 🤖 AAAI2026 (31)
🔥 Top topics: Time-Series Forecasting ×75 · Anomaly Detection ×6 · Multimodal/VLM ×4 · Diffusion Models ×3 · Adversarial Robustness ×2
- A Locally Tokenized Generative Model for Robust Time-Series Watermarking
-
The paper introduces L-VQVAE, a discrete generative model restricting token representations to bounded sliding temporal windows, and LVQMark, a watermarking scheme with alternating parity partitions, unbiased warm-up, and robust re-encoding, eliminating post-editing-induced bidirectional null drift and stabilizing clean-calibrated false-positive detection.
- A Scalable Multi-Task Model for Virtual Sensors
-
Addressing the fragmentation, lack of synergy, and linear scaling costs of conventional virtual sensors, this paper proposes the first unified multi-task model that predicts unseen signals via prototype tokens and learns attention-driven input relevance for structured pruning, delivering up to \(415\times\) speedup and \(951\times\) memory reduction.
- A Set-Sequence Model for Time Series
-
Addressing the challenge of modeling large, high-dimensional cross-sectional time series in finance and engineering, Set-Sequence explicitly decouples permutation-invariant set summarization from per-unit sequence modeling, extracting global cross-sectional latent states in linear time while supporting unaligned series and variable unit sizes at inference without manual feature engineering.
AdaST: Adaptive Coupling for Spatial-Temporal Forecasting
- AdaST: Adaptive Coupling for Spatial-Temporal Forecasting
-
AdaST decomposes spatial-temporal inputs into temporal-specific, spatial-specific, and jointly coupled representations, models them separately, and adaptively recomposes them with correlation-modulated gates, achieving the best reported results on four short-term sensor forecasting benchmarks and reducing PurpleAir MAE from the strongest comparator's 0.511 to 0.489.
- AeroChem: Closed-loop Physics-Informed State Space Modeling for Long-term Chemically-Reactive Air Quality Forecasting
-
Addressing error accumulation, physical drift, and over-smoothing of abrupt pollution spikes in long-horizon chemically-reactive air quality forecasting, AeroChem introduces a dual-branch closed-loop co-correction architecture that recursively fuses probabilistic virtual observations from S-Mamba with UDE physical priors via Neural Kalman Fusion in latent state space.
- Arbitrarily Conditioned Hierarchical Flows for Spatiotemporal Events
-
This paper introduces Arbitrarily Conditioned Hierarchical Flows (ARCH), a hierarchical flow matching framework that decomposes spatiotemporal event generation into temporal and spatial flows paired with hybrid masking, simultaneously achieving arbitrary conditional inference and closed-form, accurate instantaneous event risk (conditional intensity) evaluation.
- Benchmarking Sensor-Fault Robustness in Time Series Forecasting
-
SensorFault-Bench introduces the first cyber-physical system (CPS) sensor-fault robustness benchmark for time series forecasting, featuring 8 standardized scenarios across value, timing, and availability fault families alongside a tripartite metric protocol that decouples relative degradation from absolute error, revealing sharp performance drops in nominal-MSE-favored models and foundational forecasting models.
- Beyond Empirical Support: Structured Outlier Generation via Sinkhorn Optimal Transport
-
SBOG uses Sinkhorn support energy to identify latent boundary anchors and selects locally perturbed candidates with weak support but controlled semantics, raising CARLA's average pointwise F1 across five datasets from 0.362 to 0.432 and improving image out-of-distribution detection; these results measure downstream utility, not recovery of the true outlier distribution.
- CAST: Causal Anchored Simplex Transport for Distribution-Valued Time Series
-
To tackle latent transition-kernel aliasing where identical observed distributions evolve differently across regimes, CAST structures the transition operator directly on the probability simplex via causal successor retrieval, persistence anchoring, and bounded local stochastic transport on ordered supports, substantially curbing multi-step autoregressive rollout drift.
- COMET: Codebook-based Online-adaptive Multi-scale Embedding for Time-series Anomaly Detection
-
COMET proposes a codebook-based online-adaptive multi-scale embedding framework for time-series anomaly detection that extracts variable-wise and inter-variable features across multiple patch resolutions, quantizes embeddings into a normal pattern coreset to compute dual anomaly scores via quantization error and local density-scaled memory distance, and performs threshold-free online test-time adaptation via codebook activation pseudo-labels and contrastive learning.
- CVTA: Cross-Variable Temporal Attention for Multivariate Irregular Time Series Prediction
-
Addressing asynchronously sampled and heterogeneous multivariate irregular time series, CVTA recursively maintains a compact, interpretable state of recent observations, elapsed times, and local slopes, leveraging variable-conditioned embeddings and dual-branch cross-variable temporal attention to achieve superior short-term forecasting, \(\mathcal{O}(F)\) streaming online efficiency, and faithful interpretability.
- Cylindrical Geodesic Flow Matching for Quasiperiodic Physiological Signal Transformation
-
This paper introduces Cylindrical Geodesic Flow Matching, which embeds the instantaneous phase and positive amplitude of quasiperiodic physiological waveforms onto the cylindrical manifold \(S^1 \times \mathbb{R}_{>0}\), replaces standard affine interpolation bridges with closed-form cylindrical geodesics, and parameterizes the learned velocity field intrinsically in the tangent space, reducing HT Distance, \(L_2\) error, and DTW distance by up to \(\sim 15\%\) over the strongest competing baselines on zero-shot PPG and limited-support SCG translation benchmarks.
- Detect, Explain, Interpret: An End-to-End Benchmark for Time Series Anomaly Detection, Explainability and Interpretability
-
Introduces SHAD, a real-world high-dimensional time series anomaly detection benchmark from the Scality RING distributed storage system (215 series, 171 dimensions, with per-dimension and semantic XML labels), and provides an end-to-end evaluation spanning detection (18 baselines), explanation (feature attribution), and interpretation (frozen LLMs).
- DiffPTS: Rethinking Diffusion ELBO for Probabilistic Time Series Forecasting
-
DiffPTS jointly trains history-conditioned Gaussian mean and variance estimators with a denoising network, replacing hand-crafted variance supervision with a Gaussian negative log-likelihood derived from the diffusion ELBO and improving CRPS/MSE across nine benchmarks, including a reduction from 0.378/0.637 to 0.270/0.456 against NsDiff on Traffic, without uniformly dominating calibration metrics.
- Direct Estimation of Schrödinger Bridge Time-Series Drifts: Finite-Sample, Asymptotic, and Adaptive Guarantees
-
For time-series Schrödinger bridge (SBTS) drift estimation, this paper bypasses entropic optimal transport potentials and Sinkhorn iterations by analyzing a direct Nadaraya-Watson kernel plug-in estimator on the closed-form conditional-ratio representation, establishing uniform finite-sample bounds, pointwise asymptotic normality, and adaptive minimax rate optimality.
- Dirichlet-Guided Group Forecasting for Alleviating Over-smoothing in Time Series Forecasting
-
Addressing the pervasive over-smoothing problem in multi-step time series forecasting caused by non-convex mode compression under single-realization supervision, this paper proposes DGF, which explicitly models mode-conditioned predictive distributions and Dirichlet second-order uncertainty, optimized with hierarchical GRPO over accuracy, dynamic consistency, and diversity rewards.
- Discretizing Continuous Time Series for Imputation with Masked Diffusion Training
-
Addressing representation entanglement between missing and observed coordinates and indirect noise-prediction objectives in continuous diffusion, MDTIM adapts masked discrete diffusion to continuous time series through stochastic discretization, ordinal-aware soft labeling, and expectation decoding.
- Does Synthetic Data Help? Empirical Evidence from Deep Learning Time Series Forecasters
-
Through a comprehensive empirical study of 9 experiment groups and 4,218 runs, this paper demonstrates that synthetic time series augmentation is sharply architecture-conditional: channel-mixing models (TimesNet, iTransformer) systematically benefit and can even surpass full-data baselines in low-resource settings, whereas channel-independent models (DLinear, PatchTST) are consistently degraded.
- Dynamic Dual-Feedback Conformal Inference for Time Series Forecasting
-
To alleviate severe interval oscillation caused by high-variance stochastic noise in online non-exchangeable time series conformal prediction, this paper introduces Dynamic Dual-feedback Conformal Inference (DDCI), which couples actual error feedback with empirical conformal threshold feedback in a push-pull mechanism to produce substantially tighter and smoother prediction intervals while strictly guaranteeing nominal coverage.
- E4GEN: Event-level Explainable Extreme-Enhanced Time-series Generation
-
E4GEN introduces an explainable diffusion framework for event-level extreme-aware time-series generation that systematically addresses when, what, and how to control extreme-event synthesis through a dataset-adaptive activation step (E-Activator), self-driven semantic prediction with a data-conditioned training mechanism (E-Predictor with DCT-NIS), and layer-aligned gated residual injection (E-Control).
- Existence Precedes Value: Joint Modeling of Observational Existence and Evolving Values in Irregular Time Series
-
Addressing the unrealistic oracle assumption in irregular time series models that future observation timestamps are known a priori, Timeflies introduces a unified dual-stream framework to jointly model future observational existence and evolving state values, validated on the Shadow benchmark with the novel OVJE metric.
- Existence Precedes Value: Joint Modeling of Observational Existence and Evolving States in Time Series Forecasting
-
Challenging the unrealistic oracle assumption that future valid timestamps are known in advance, this paper proposes Timeflies, a dual-stream framework that reformulates irregular time series forecasting into joint prediction of future observational existence and numerical states, substantially outperforming prior SOTAs on the 31-dataset Shadow benchmark.
- Explanation of Dynamic Physical Field Predictions Using WassersteinGrad: Application to Storm Surges
-
To tackle spatial geometric displacements and phase errors induced by gradient smoothing in autoregressive physical forecasting models, this paper introduces WassersteinGrad, which extracts a geometric consensus of perturbed attribution maps via entropic 2-Wasserstein barycenters, delivering markedly improved faithfulness and structural robustness.
- Faithful Embeddings of Irregular and Asynchronous Data for Online Log-NCDEs
-
For irregular and asynchronous observational data, this paper introduces a continuous and injective faithful embedding that builds interval log-signatures directly from increments without explicit interpolation, driving online Log-NCDEs with up to 3000x training speedups and high robustness against extreme input dropping.
- Falcon-X: A Time Series Foundation Model for Heterogeneous Multivariate Modeling
-
Falcon-X decouples physical variates into a unified latent prototype space using differential attention to capture both synergistic and antagonistic affinities, achieving state-of-the-art zero-shot multivariate forecasting on GIFT-Eval and fev-bench.
- FAME: Forecasting Academic Impact via Continuous-Time Manifold Evolution
-
To overcome the limitations of static LLMs that evaluate scientific papers in a temporal vacuum, FAME models research evolution as continuous-time latent manifolds, achieving robust prospective impact forecasting via geometric alignment between paper residuals and dynamic topic momentum.
- FAST-Brain: A Flow-Aligned Spatio-Temporal Surrogate Brain Model
-
Addressing the challenges of long-horizon temporal dependencies, anatomical spatial priors, and high-dimensional ambient noise, FAST-Brain proposes a flow-aligned spatio-temporal surrogate brain model that directly predicts clean BOLD signals, proving that approximation errors scale with intrinsic subspace dimension rather than ambient ROI dimension and achieving state-of-the-art functional and effective connectivity recovery.
- FLAME: Flow Enhanced Legendre Memory Models for General Time Series Forecasting
-
FLAME integrates translated and scaled Legendre polynomial memory units with a masked autoregressive normalizing flow head, delivering an ultra-lightweight (2M to 10M parameters) foundation model that excels at both deterministic and probabilistic zero-shot time series forecasting.
- From Jumps to Signatures: A Generative Method for Temporal Point Processes
-
Addressing the limitation that neural temporal point processes (TPPs) optimize per-event local likelihoods and cannot natively apply rough path signatures to càdlàg jump paths, this paper introduces the stable, injective interarrival embedding to continuous bounded-variation paths, derives rigorous counting-path distributional discrepancies, and develops SigTPP, a generative TPP model trained via a global signature Wasserstein loss on complete trajectories.
- From Retrieval to Recognition: Knowledge-Enhanced Post-training of Time Series Foundation Models for Classification
-
This paper introduces Retrieval-to-Recognition (R2R), a plug-and-play post-training paradigm that extracts and retrieves discriminative numerical patterns from a pretraining knowledge base using a gradient-free query encoder, adaptively fusing them with TSFM parameter representations to consistently outperform standard fine-tuning across 158 datasets.
- GlucoFM: A Dual-Stream Foundation Model for Continuous Glucose Monitoring
-
GlucoFM is a lightweight self-supervised foundation model for continuous glucose monitoring (CGM) that preserves observation masks on a 24-hour chronological grid, decomposes glucose dynamics into slow glycemic trends and short-term deviations via learnable causal filtering, and pretrains using dual JEPA-style latent objectives, outperforming existing general time-series and CGM-specific models across metabolic phenotyping and postprandial response forecasting.
- GlucoFM-Bench: Benchmarking Time-Series Foundation Models for Blood Glucose Forecasting
-
GlucoFM-Bench establishes the first standardized benchmark evaluating eight time-series foundation models, time-series LLMs, and supervised architectures across 15 real-world CGM datasets, revealing that pre-trained foundation models deliver competitive zero-shot transfer—especially in critical hypoglycemic ranges—while lightweight task-specific LSTMs still set the upper bound under data-abundant supervision.
- Hermes: A Multi-Scale Spatial-Temporal Hypergraph Network for Stock Time Series Forecasting
-
Hermes presents a multi-scale spatial-temporal hypergraph network for stock forecasting that models inter-industry lead-lag relationships via patch-based moving aggregation and aligns multi-frequency features using learnable Mahalanobis distance metric learning, achieving state-of-the-art predictive accuracy and Sharpe ratios across major US stock benchmarks.
- High-Dimensional Gaussian Graphical Model Testing for Long-Memory Time Series
-
This paper develops a data-adaptive testing procedure for conditional independence graph structures in high-dimensional stationary Gaussian time series with long-range dependence, using off-support local-inverse residuals and an overlapping block bootstrap to establish finite-sample Gaussian approximation bounds and two-sample differential graph testing without estimating the full precision matrix.
- In-Context Learning to Predict Critical Transitions in Dynamical Systems
-
TipPFN is the first Prior-Data Fitted Network (PFN) for anticipating critical transitions in dynamical systems, pretrained on a vast synthetic prior of stochastic bifurcations to deliver state-of-the-art zero-shot and in-context early warnings via a single forward pass.
Kairos: Toward Adaptive and Parameter-Efficient Time Series Foundation Models
- Kairos: Toward Adaptive and Parameter-Efficient Time Series Foundation Models
-
Kairos handles local time-series complexity and instance-level spectral differences through mixture-of-size encoding and dynamic rotary positional encoding, then predicts multiple future patches in parallel, reaching a normalized MASE of 0.738 on GIFT-Eval with 53M parameters, without uniformly leading in probabilistic or task-level performance.
- KITE: A Knowledge-Injected Temporal Encoder for Time-Series Event Forecasting
-
Addressing the extreme scarcity of precursor signals for rare, high-cost events in multivariate time series, KITE injects domain engineering knowledge (periodicity, conservation laws) via switchable, zero-initialized covariate and attention-bias mechanisms into a self-supervised JEPA backbone, substantially boosting event forecasting under short contexts and out-of-domain transfer.
- Learning the Context of Errors: Black-Box Online Adaptation of Time Series Foundation Models
-
Addressing the challenge where commercialized time series foundation models (TSFMs) are deployed as closed-source APIs with no gradient or weight access, this paper presents ORCA, a black-box online adaptation framework that learns error distributions conditioned jointly on history inputs and base predictions, incorporating a structural-bias linear adapter, decay replay buffer, and a Boltzmann confidence router.
- LogSig-SSM: Time-Series Modelling with Multi-Scale Log-Signature Compression for State-Space Models
-
Addressing the excessive GPU memory overhead of selective state-space models (SSMs) on ultra-long sequences and their single-block inability to capture higher-order cross-channel path geometry, LogSig-SSM tokenises multivariate time series via multi-scale windowed log-signatures before recurrent propagation, achieving top-tier accuracy, up to 37× memory reduction, and intrinsic robustness to irregular sampling.
- Massively Parallel Exact Inference for Hawkes Processes
-
By reformulating the recursive intensity and exact parameter gradients of linear exponential Hawkes processes into associative products of sparse augmented transition matrices, this paper leverages the Blelloch parallel prefix scan alongside a constant-memory batching scheme to enable exact maximum likelihood estimation scaling to tens of millions of events on modern GPUs.
- MissPath-FM: Flow Matching with Structured Priors for Partially Observed Time Series
-
MissPath-FM tackles the transport inefficiency of conditional flow matching on incomplete time series—where starting from uninformed Gaussian noise wastes capacity on already-observed entries—by coupling structured source priors with a missingness-aware smooth probability path, winning 45 out of 48 evaluation settings against eight baselines.
- MMTA: Benchmarking Multimodal Temporal Analysis with Time Series, Text, and Vision
-
MMTA introduces the first real-world tri-modal benchmark pairing time series, text, and scene vision across nine domains and two cross-modal layouts, revealing that textualizing time series triggers severe context explosion and temporal misalignment in current MLLMs, which the proposed TimeOmni-v resolves using a dedicated Chronos-2 patch encoder and timeline-aligned multimodal RoPE.
- MoCAR: Motion-code Coordinate-aware AutoRegression for Continuous Trajectory Forecasting
-
MoCAR learns endpoint-normalized continuous motion codes to cast trajectory forecasting as next-code autoregression in latent space, achieving top-tier multimodal forecasting and strong zero-shot transfer without trajectory-space re-tokenization or proposal-and-refinement pipelines.
- MoCAR: Motion-code Coordinate-aware AutoRegression for Continuous Trajectory Forecasting
-
MoCAR learns endpoint-normalized continuous motion codes to cast trajectory forecasting as next-code autoregression in latent space, achieving top-tier multimodal forecasting and strong zero-shot transfer without trajectory-space re-tokenization or proposal-and-refinement pipelines.
- MoveBench: A Benchmark for Global-Scale Wildlife Movement Forecasting
-
MoveBench establishes the first global-scale benchmark for probabilistic animal movement forecasting across 110 wildlife species paired with 1.6 billion environmental raster tiles, revealing severe out-of-distribution generalization bottlenecks across individuals for deep learning models relative to classical ecological baselines.
- Multivariate Time Series Forecasting needs Cross Variable Loss
-
CvLoss leaves the forecasting backbone unchanged and supplements point-wise MSE with constraints on residual differences between forecast patches from different variables, achieving 111 wins, 2 ties, and 1 loss across the paper's 114 backbone-controlled comparison cells, with no loss computation at inference time.
- Njord: A Probabilistic Graph Neural Network for Ensemble Ocean Forecasting
-
Njord introduces the first conditional deep latent variable hierarchical graph neural network for high-resolution global and regional ocean forecasting, combining spherical K-means geometry-adaptive meshes with sea ice physical constraints to efficiently sample calibrated ocean ensemble forecasts in a single forward pass.
- OATS: Online Data Augmentation for Time Series Foundation Models
-
OATS introduces an online data augmentation paradigm for time series foundation models, leveraging first-order Taylor data attribution scores and SNR filtering to construct Time-Series Influence Scores (TSIS) that guide a conditional diffusion model to dynamically synthesize high-value training samples, integrated with an explore-exploit scheduler that drastically cuts runtime overhead while outperforming regular training and static augmentation across six benchmarks.
- Observable Neural ODEs for Identifiable Causal Forecasting in Continuous Time
-
By bridging control-theoretic observability with causal inference, this paper proves that observable latent states serve as valid causal mediators satisfying the continuous-time conditional front-door criterion, and introduces Observable Neural ODEs (ObsNODEs) with an architectural observable normal form to achieve identifiable dynamic counterfactual forecasting under hidden confounding.
- Observable Neural ODEs for Identifiable Causal Forecasting in Continuous Time
-
By bridging control-theoretic observability with causal inference, this paper proves that observable latent states serve as valid causal mediators satisfying the continuous-time conditional front-door criterion, and introduces Observable Neural ODEs (ObsNODEs) with an architectural observable normal form to achieve identifiable dynamic counterfactual forecasting under hidden confounding.
- OliO: ODE-based Linear Transition Operator for Self-Supervised Time Series Forecasting
-
OliO introduces an ODE-derived linear transition operator for self-supervised time series pre-training, enforcing a strictly upper triangular generator that yields finite-polynomial bidirectional temporal transport to prevent exponential divergence and significantly improve long-term forecasting accuracy.
- OliO: ODE-based Linear Transition Operator for Self-Supervised Time Series Forecasting
-
Addressing the failure of existing self-supervised time series models to account for signed temporal lag and preserve continuous dynamics, OliO introduces a plug-and-play linear ODE-derived bidirectional transition operator with a strictly upper-triangular nilpotent generator that guarantees stable finite-polynomial transitions without altering downstream forecasting architectures.
- Ordinary Least Squares as an Attention Mechanism
-
This paper proves that ordinary least squares (OLS) out-of-sample predictions are algebraically equivalent to a restricted linear attention module, viewing least squares as inner-product similarity retrieval in an orthonormal predictor space, and introduces a lightweight polynomial Regression Block that matches FT-Transformer on tabular data with ~30% of the parameters.
- Otter Weather: Skillful and Computationally Efficient Medium-Range Weather Forecasting
-
Otter Weather strips away bespoke Earth-specific geometric inductive biases in favor of a standard 2D Swin U-Net augmented with 2D RoPE, SwiGLU, and the Muon optimizer, outperforming numerical weather prediction baselines on a single GPU in under 3.5 A100-days and advancing the skill-compute Pareto frontier for both deterministic and probabilistic forecasting.
- Paloma: Phase-Conditioned Residual Modulation for Time Series Forecasting
-
Addressing the common assumption that residuals in periodicity decomposition are unstructured noise, Paloma introduces phase-conditioned rotation and amplitude scaling modules to explicitly model directional and magnitude residual deviations, achieving SOTA performance and serving as a general-purpose plugin.
- PAMod: Modeling Cyclical Shifts via Phase-Amplitude Modulation for Non-stationary Time Series Forecasting
-
Addressing the non-stationary challenge of mean and variance shifts over periodic cycles in real-world time series, PAMod draws inspiration from communication modulation theory to perform additive phase modulation (mean adjustment) and multiplicative amplitude modulation (variance adaptation) in normalized feature space, achieving state-of-the-art accuracy across 12 benchmarks with 5–10× fewer parameters than Transformer baselines.
- Parallel-in-Time Training of Recurrent Neural Networks for Dynamical Systems Reconstruction
-
Addressing the linear computational bottleneck of backpropagation through time (BPTT) and chaos-induced gradient explosions, this paper integrates Generalized Teacher Forcing (GTF) into the parallel root-finding framework DEER, establishing the GTF-DEER algorithm with provable contraction convergence to achieve up to an \(870\times\) speedup and enable high-fidelity reconstruction on ultra-long trajectories (\(T > 10^4\)).
- PG-LRF: Physiology-Guided Latent Rectified Flow for Electro-Hemodynamic PPG-to-ECG Generation
-
PG-LRF incorporates a coupled electro-hemodynamic ODE simulator into a latent rectified flow model, utilizing a physiology-aware autoencoder and terminal Euler residual guidance to reconstruct diagnostically faithful ECG waveforms from wearable PPG signals.
- PrismFlow: Residual Dynamics for Flow Matching in Time-Series Generation
-
PrismFlow mitigates the velocity averaging and mode collapse problem in flow matching for time series by augmenting a stable global transport field with a bank of dissipativity-constrained Koopman latent linear dynamical experts routed via confidence-aware Winner-Take-All selection.
- Progressive Memory Transformer: Memory-Aware Attention for Time-Series
-
PMT introduces Progressive Memory Transformer, which equips sliding windows with writable, window-aligned memory slots that expose mid-range motifs as first-class representations alongside tokens and sequence summaries, achieving state-of-the-art low-label classification across seven benchmarks (1% labels beating prior 5% baselines) and superior long-horizon forecasting under a three-scale contrastive learning framework.
- Progressive Memory Transformer: Memory-Aware Attention for Time-Series
-
PMT exposes writable memory updated along sliding windows as an explicit mid-range representation and separately supervises tokens, memory, and sequence summaries, achieving 84.4% average accuracy across seven low-label classification datasets while a cue-retention probe verifies that memory carries information across windows.
- Raw-Routed Mixture of Adapters: A Causal Intervention for Routing Collapse in Time Series Foundation Models
-
Addressing the "normalization-induced routing collapse" that occurs when adapting instance-normalized time series foundation models with mixture-of-experts, this work proposes Raw-Routed Mixture of Adapters (RR-MoA)—routing on pre-normalization raw inputs to enable efficient, high-performing adaptation with a strictly frozen backbone.
- Raw-Routed Mixture of Adapters: A Causal Intervention for Routing Collapse in Time Series Foundation Models
-
RR-MoA gives a lightweight gate the input before instance normalization while adapter experts retain frozen-backbone hidden states, repairing an input-side information bottleneck and winning all 54 paired comparisons reported against fixed adapters, although the basic version still trails a linear predictor trained from scratch.
- Realism VS Accuracy: Event Sequence Forecasting from a Generative Modeling Perspective
-
This paper frames multi-horizon forecasting of event sequences with heterogeneous features as a conditional generative modeling task, exposes a fundamental Pareto trade-off between prediction accuracy and distributional realism, and introduces Wasserstein BaryBooster to aggregate stochastic diffusion rollouts into state-of-the-art deterministic forecasts.
- ResCast-100K: A Comprehensive Dataset for Cross-Domain Residential Load and Indoor Temperature Forecasting
-
ResCast-100K tackles the lack of cross-domain benchmarks and coupled-target formulations in residential forecasting by providing 15-minute time series for 100k EnergyPlus-simulated homes alongside 5 standardized real-world datasets, demonstrating that MLP-Mixers and cross-attention architectures significantly outperform classical models under structured domain shift and sim-to-real transfer.
- Rethinking Time Series Tokenization from a Frequency Perspective
-
Addressing spectral leakage and aliasing distortions in patch and phase tokenizers, this paper establishes distortion-free boundary conditions and proposes TimeFT—a parameter-free, plug-and-play tokenizer yielding significant gains across forecasting, classification, and anomaly detection.
- Retrieval-Augmented Diffusion Modeling for Stochastic Discount Factor Portfolios
-
Addressing the non-stationary regime shifts and noisy multimodal inputs that distort the stochastic discount factor (SDF) pricing manifold, RADAR introduces a retrieval-augmented conditional diffusion framework that parameterizes state-dependent noise distributions from historically similar regimes, achieving state-of-the-art risk-adjusted returns and economically meaningful asset co-movements.
- Retrieval Over Training: Similarity search-based Model Selection for Time Series Anomaly Detection
-
Addressing the limitations of train-dependent meta-selectors and costly full-pool ensembling in time series anomaly detection, RAMSAD introduces a training-free retrieval framework that leverages TSFM segment embeddings and multi-positive multimodal semantic alignment to transfer precomputed detector performance profiles from nearest-neighbor series, outperforming trained selectors in ID and matching full-pool ensembles in OOD.
- Revitalizing Medical Time Series with Vision-Informed Retrieval: A Vision-Language Perspective
-
Addressing the limitation of existing medical time series models that treat signals purely as numerical sequences while ignoring visual waveform morphology, ViRe extracts morphology-aware Vision Queries from rendered waveform plots using a frozen CLIP model to retrieve clinical evidence from numerical features via cross-attention, achieving an average 6.42% relative improvement across six public benchmarks.
- Revitalizing Medical Time Series with Vision-Informed Retrieval: A Vision-Language Perspective
-
ViRe renders the same numerical EEG/ECG segment as a waveform and uses frozen CLIP visual features as a query to retrieve evidence from temporal and channel numerical tokens; the mean six-metric score across six benchmarks rises from Medformer's 77.90 to 82.90, but ViRe does not lead on every dataset, and this is not patient-level diagnostic accuracy.
- SCDM: Scalable Causal Discovery in Nonlinear Temporal Systems with Meta-Learning
-
SCDM formulates nonlinear temporal causal discovery as shared-parameter episodic meta-learning across target-variable tasks, combining target-task subsampling and temporary adaptation to update a global causal-strength matrix, and extracts continuous directed graphs via post-training perturbation or direct structural readouts with remarkable scalability on systems up to 3,000 variables.
- SDFlow: Similarity-Driven Flow Matching for Time Series Generation
-
Addressing the cascading error accumulation from exposure bias and high-dimensional transport challenges in autoregressive discrete time-series generation, SDFlow introduces a non-autoregressive flow matching framework operating entirely within a frozen VQ-VAE latent space. By coupling an end-to-end low-rank manifold decomposition with a topology-preserving anchor prior and categorical posterior supervision, SDFlow delivers state-of-the-art long-horizon fidelity alongside a 3–10× inference speedup.
- SeoulMMOD: A Large-Scale Multimodal Origin-Destination Flow Benchmark
-
SeoulMMOD introduces the first large-scale multimodal origin-destination flow forecasting benchmark covering 3 years (2023–2025), 6 urban transit modes, and two spatial scales (25 districts and 426 sub-districts) in Seoul, benchmarking 14 forecasting architectures and identifying key scalability and joint-training bottlenecks.
- SPATIALEPIBENCH: Benchmarking Spatial Information and Epidemic Priors in Forecasting
-
Presents SpatialEpiBench, a comprehensive spatiotemporal forecasting benchmark across 11 multivariant epidemic datasets, demonstrating that adjacency-informed graph models consistently underperform a naive persistence baseline even during outbreaks, and diagnosing the root causes of spatial modeling failures.
- STAR: boosting time series foundation models for anomaly detection through STate-aware AdapteR
-
STAR introduces a plug-and-play STate-aware AdapteR for Time Series Foundation Models, combining an identity-guided state memory encoder, dynamic conditional bottleneck adaptation, and patch-level numeral-state matching to unlock discrete state modeling and boost multivariate anomaly detection across cyber-physical systems.
- SteerCast: Retrieval-Based Latent Steering for Decoder-Only Time Series Forecasting
-
SteerCast addresses cumulative error propagation in decoder-only autoregressive time series forecasters by introducing an inference-time latent steering mechanism that retrieves training-set histories, derives model-aware residual vectors between ground-truth and predicted continuations, and injects them into hidden states via adaptive cosine gating without parameter updates.
- STEP: Learning STructured Embeddings for Progressive Time Series
-
STEP constructs a structured low-dimensional manifold anchored between orthogonal prototype vectors via cosine triplet contrastive learning, extracting a polar "latent compass" \((\theta, r)\) that tracks irreversible degradation and operating modes without proxy labels—enabling a simple linear regressor to match or surpass deep neural architectures on RUL estimation.
- TempoPFN: Synthetic Pre-training of Linear RNNs for Zero-shot Time Series Forecasting
-
TempoPFN presents the first univariate time series foundation model built on linear Recurrent Neural Networks with GatedDeltaProduct recurrence, pre-trained exclusively on an open-source synthetic pipeline spanning regime-switching Ornstein–Uhlenbeck SDEs, causal DAGs, and procedural audio synthesis; armed with an inter-layer state-weaving mechanism, it eliminates windowing heuristics, delivers fully parallelized training alongside constant-time quantile forecasting, and outperforms all synthetic baselines while rivaling top real-data models on Gift-Eval and fev-bench.
- TENET: Time-point Encoding Network for Multivariate Time Series Anomaly Detection
-
Reframes multivariate time series anomaly detection as estimating the compatibility of a current-point system state with normal behavior, proposing TENET to actively retrieve temporal context anchored at the target point and scoring anomalies via a fixed standard-normal negative log-likelihood.
- TimeClaw: A Time-Series AI Agent with Exploratory Execution Learning
-
TimeClaw addresses the dual bottlenecks of completion-only supervision and tool-prior collapse in time series agents via an "Explore-Compare-Distill-Reinject" loop, leveraging verifiable metric supervision and hierarchical external experience to outperform strong baselines across 17 finance and weather tasks while keeping the base LLM frozen.
- TimeES: Probabilistic and Deterministic Time Series Forecasting via Evolutionary Spectra
-
TimeES predicts time-varying complex spectral amplitudes with a lightweight network and synthesizes trajectories using fixed Fourier phases and frequency-level random variables shared across the forecast horizon, unifying deterministic and probabilistic forecasting; its average performance on nine probabilistic benchmarks is strong, but ILI clearly deteriorates, and spectral recovery and uncertainty training require careful interpretation.
- TimeTok: Granularity-Controllable Time-Series Generation via Hierarchical Tokenization
-
TimeTok introduces a unified framework for Granularity-Controllable Time-Series Generation (GC-TSG) by encoding time series into coarse-to-fine discrete tokens, combining flow-matching decoding with block-wise autoregressive modeling to achieve both unconditional generation and conditional refinement from coarse sketches.
- TimeTok: Granularity-Controllable Time-Series Generation via Hierarchical Tokenization
-
TimeTok encodes time series into discrete token prefixes with coarse-to-fine semantics, adds detail through granularity-block autoregression, and controls output granularity through conditional flow matching, achieving strong conditional refinement and distributional metrics without leading on every downstream predictive metric.
- TiRex-2: Generalizing TiRex to Multivariate Data and Streaming
-
TiRex-2 extends the recurrent xLSTM-based time series foundation model with bidirectional time mixing and asymmetric grouped-attention variate mixing, allowing future-known covariates to be exploited without leaking future target information, and enabling constant-cost streaming updates over arbitrarily long horizons.
- To See Far, Look Close: Evolutionary Forecasting for Long-term Time Series
-
Addressing the rigid coupling and gradient conflict pathology of Direct Forecasting (DF), this paper proposes Evolutionary Forecasting (EF), which decouples the output horizon from evaluation horizons via block-wise reasoning and multi-phase context evolution, outperforming task-specific DF models across all horizons with a single training session.
- Topological Periodicity Test (TopPT) via Confidence Bound of Time-Delay Embeddings
-
TopPT turns an early-born, long-lived loop in a delay point cloud into a statistical test with a subsampling confidence radius, providing asymptotic error control over restricted, quantitatively separated signal classes; in synthetic experiments, the raw rule detects both periodic signals without rejecting four structured non-periodic signals, but interpolation-error correction substantially reduces detection, and real respiratory data only assess local loop detection.
- TS-ICL: A Flexible Time-Indexed Foundation Model for Time Series via In-Context Learning
-
TS-ICL reformulates time series forecasting and imputation as continuous timestamp-aligned in-context regression, combining a DAG-based causal synthetic prior with an encoder–regressor architecture to achieve state-of-the-art zero-shot imputation and robust covariate-aware forecasting.
- TSQAgent: Rating Time Series Data Quality via Dedicated Agentic Reasoning
-
To tackle the poor dimension generalization and weak numerical reasoning of LLMs in time series data quality assessment, this paper proposes TSQAgent, a multi-agent framework featuring a Perceiver, Inspector, and Adjudicator that leverages tailored GRPO for adaptive dimension planning and 16 external analytical tools for grounded pairwise rating and data selection.
- TStruct: Learning Shared Temporal Structures for Long-term Time-Series Forecasting
-
To resolve the vulnerability of deep models overfitting incidental temporal interactions in noisy and non-stationary time series, TStruct directly parameterizes a compact set of globally shared temporal dependency structures and performs adaptive soft allocation using dynamic sample features and static channel priors, consistently outperforming leading Transformer and linear baselines across 7 major benchmarks while maintaining high efficiency.
- Tyche: One Step Flow for Efficient Probabilistic Weather Forecasting
-
To tackle the severe inference bottleneck of multi-step generative weather forecasters, Tyche models the conditional predictive distribution via a destination-aware average-velocity flow with Jacobian-vector product (JVP) rectification for distillation-free 1-NFE sampling, combined with curriculum CRPS rollouts to mitigate autoregressive error accumulation, matching multi-step diffusion quality while running 12x–45x faster.
- VACE: Learning Geometrically Structured Representations for Time Series Anomaly Detection
-
VACE dispenses with negative pair-sampling heuristics in contrastive learning, using depthwise-separable convolutions to preserve sensor-specific features and a forward-backward velocity consistency objective to shape normal time series into compact, directionally coherent latent trajectories, achieving state-of-the-art anomaly detection on TSB-AD-M with minimal computational budget via multiplicative positional and directional scoring.
- Why Do Time Series Models Need Long Context Windows?
-
Reveals that the reliance of global time series models on long observation windows is driven by generative process identification (GPI) rather than solely capturing long-range temporal dependencies, and proposes a decoupled architecture that amortizes GPI inference costs to deliver Pareto-optimal forecasts with up to 100× acceleration.
- Memory Determines Learning Direction: A Theory of Gradient-Based Optimization in State Space Models
-
This paper formalizes the gradient optimization dynamics of linear state space models (SSMs) through the memory function framework, proving that the Vandermonde projection operator of recurrent eigenvalues strictly dictates the backpropagation update direction, theoretically establishing the equivalence between S4 and diagonal S4D, and demonstrating that fixing recurrent weights in a reservoir computing scheme prevents overfitting while accelerating convergence.