Time-Series Foundation Models

Overview

Time-series foundation models learn reusable patterns from large collections of sequences, aiming to support tasks such as forecasting, classification, imputation, and anomaly detection across domains. Common approaches include Transformer models with patch or value tokenization, masked and contrastive pretraining, next-token prediction, and newer recurrent or state-space architectures; training data may mix real series with synthetic ones. Models are used zero-shot or adapted through fine-tuning, in-context examples, covariates, retrieval, and lightweight adapters, with benchmarks such as GIFT-Eval and UCR/UEA testing different capabilities. Transfer can reduce the need for task-specific training, but results vary with data and operating conditions: specialists may still win, and robustness to distribution shifts, long contexts, uncertainty, and fair evaluation remain open challenges.

Latest papers

345 papers54 papers in the last four weeks, down 17% on the four weeks before. 0.4% of all new papers.

Nov 1, 2026

SensOFormer: Foundation Models for Particle Accelerators

Large-scale scientific facilities, such as particle accelerators, are complex systems composed of thousands of interacting components that must operate collectively to support diagnostics and control, and ensure safe, reliable operation. These facilities rely on extensive sensor networks that generate large volumes of heterogeneous, incomplete, and noisy telemetry, often sampled at different rates across subsystems. In this paper, we adopt the foundation-model paradigm to learn general-purpose sensor representations from historical accelerator data through self-supervised pretraining. We introduce SensOFormer, a denoising masked transformer with a Perceiver-style encoder-decoder architecture designed to handle variable sensor sets, accommodate heterogeneous sampling rates, and learn robust representations from noisy measurements. The model is pretrained on multiple particle-accelerator datasets and transferred to downstream tasks, including missing-data imputation, anomaly detection, cavity identification, and fault identification. Through extensive evaluations and comparisons, we demonstrate that a single self-supervised model can learn reusable representations for diverse operational tasks in large-scale scientific facilities. Project page: https://mahindrautela.github.io/sensoformer.
Nov 1, 2026

GaitFM: A Cross-Modal Foundation Model for Wearable Gait Assessment

Gait assessment supports the diagnosis, monitoring, and prognosis of diverse movement-related conditions, such as neurological and musculoskeletal disorders. Recent advances in deep learning using everyday wearable inertial measurement units (IMUs) offer an accessible and low-cost approach to automated gait assessment. However, IMU measurements vary substantially across individuals, body locations, and device configurations, while existing deep learning models are typically developed using small, homogeneous datasets for specific clinical tasks, limiting their robustness and generalizability in diverse real-world contexts. Here we present GaitFM, a cross-modal foundation model for gait assessment using IMU sensors. GaitFM improves the generalizability of gait assessment by transferring knowledge from large vision models and pretraining on unlabeled IMU sensor data. We pretrain GaitFM on a large curated dataset comprising about 4,000 participants and 1.23 million gait windows across diverse clinical and daily conditions, sensor modalities, and body locations. GaitFM demonstrates strong performance across 18 gait-related tasks, including 3 gait parameter estimation tasks, 14 disease and gait abnormality detection tasks, and 1 clinical outcome prediction task. Furthermore, GaitFM generalizes across different sensor configurations and body locations, including those commonly used in clinical settings and everyday wearable devices. These findings demonstrate the effectiveness of transferring visual knowledge in improving sensor-based gait assessment performance, offering a new promising pathway for addressing data scarcity in the sensor and medical domains.
Oct 30, 2026

Chronos-2: A foundation model for multivariate time series forecasting

Accurate forecasting with uncertainty quantification is critical in a variety of scientific applications ranging from epidemiology and economics to energy systems and earth sciences. Yet most forecasting models still need to be developed and trained separately for each domain-specific dataset. Time series foundation models (TSFMs) aim to remove this limitation and enable the application of the same model across domains through large-scale pretraining. However, most existing TSFMs treat each time series independently and therefore cannot model interactions among time series and external drivers. Here, we introduce Chronos-2, a foundation model for zero-shot forecasting that models the interactions between multivariate time series and heterogeneous covariates, which commonly arise in scientific applications. The model infers relationships among time series directly from historical context and generates probabilistic forecasts without task-specific treatment. Across various benchmarks, Chronos-2 achieves state-of-the-art accuracy on average by improving skill scores by 5.2 percentage points relative to the strongest baselines and increasing win rates by 10 percentage points. Chronos-2 is also computationally efficient, generating forecasts for 100 time series in under one second on a single mid-range GPU, 50x and 1000x faster end-to-end than the best task-specific point and probabilistic forecasters, respectively. In addition, we demonstrate its effectiveness in diverse scientific settings -- including infectious-disease forecasting, macroeconomic modeling, carbon-flux prediction, and runoff simulation -- where the model matches or surpasses specialized methods developed for each domain. These results show that a single pretrained model can capture structure shared across temporal systems.
Oct 30, 2026

QUAL: Quantile Alignment Improves Zero-Shot Time-Series Foundation Models

Time-series foundation models (TSFMs) provide zero-shot forecasts for unseen series, but their accuracy varies across datasets. Combining several models can reduce this variability, yet zero-shot deployment provides no target labels for selecting or weighting them. We find that a model pool is useful not only as a final ensemble, but also as a probability reference for transforming each member forecast. Aligning quantiles through this pool-relative reference improves every member forecast in aggregate. Motivated by this finding, we propose QUAL (Quantile alignment), a simple label-free framework for improving and combining diverse TSFMs. QUAL first produces aligned member forecasts, then applies Huber aggregation and rearrangement to obtain a robust combined forecast without training or model modification. On GIFT-Eval, quantile alignment reduces average member MASE and CRPS by 7.4% and 9.8%. The full framework remains robust when weak forecasters enter a candidate pool whose quality cannot be verified in advance. Code is available at https://github.com/junhyeokkang/QUAL.
Oct 29, 2026

Understanding axial attention in TSFMs through single location regression

Recent advances in Time Series Foundation Models (TSFMs) increasingly rely on axial attention mechanisms, but despite their empirical success, the theoretical properties of axial attention remain largely unexplored. To bridge this gap, we extend the Single Location Regression framework introduced in Marion et al. [2025] to 2D structured data, in order to study sparse information settings where a target depends on a small number of tokens. We formulate an analytically tractable axial attention-type predictor and rigorously analyze its statistical properties and training dynamics. We prove that under a specific asymptotic regime and with knowledge of the underlying oracle parameters, our predictor achieves Bayes optimality, whereas a class of semi-linear simplifications strictly fails to do so, underlining the role of the inner non-linearity. We prove that for a specific temperature scaling, Projected Gradient Descent converges globally to the optimal oracle parameters. Extensive numerical experiments validate our theoretical findings and highlight the critical role of inverse temperature scheduling in empirical convergence.
Oct 28, 2026

Modular Bayesian Foundation Models: Bayesian time series encoders for OOD detection, routing and forecasting

Industrial systems continuously generate time series that vary in sampling frequency, scale, dynamics, and operating regime. Consequently a single time-series foundation model may not be able to represent all regimes effectively and retraining a large model for every new dataset or regime shift is computationally expensive. Modular foundation models address this issue by reusing specialised encoders trained on diverse datasets and by allowing the expansion of capabilities through the addition of new experts. This approach requires compact experts to minimize storage cost and a proper selection mechanism of experts for unseen data. We propose a modular foundation model framework that combines end-to-end Bayesian encoder training with uncertainty based routing. Using Singular Bayesian Neural Networks (SBNNs), we build a collection of compact Bayesian encoders whose epistemic uncertainty guides label-free selection. At the selected ranks, our Bayesian collection uses 63.8% less storage than its deterministic counterpart. We evaluate this framework on time-series data through distribution shift, encoder selection, and probabilistic forecasting. Across seven incoming datasets, uncertainty-based routing is competitive with distance-based routing in terms of forecast MSE, while reducing average ECE from 0.086 to 0.072 and average NLL from 0.907 to 0.836. It also increases PICP90 from 82.28% to 89.03%. In a separate model-family comparison, we evaluate the SBNN pipeline against full-rank Bayes by Backprop, deep ensembles, and MC Dropout. The SBNN pipeline achieves an average MSE of 0.390, comparable with 0.396 for deep ensembles, 0.389 for MC dropout, and 0.443 for the full-rank Bayesian model. It uses 63.9% fewer parameters than MC dropout, 81.9% fewer than the full-rank model, and 92.8% fewer than the deep ensemble.
Oct 27, 2026

Are Seismic Foundation Models More Than Glorified Feature Extractors?

Earthquake sequences are stochastic on two separated timescales: the seismic waves of any individual earthquake propagate over seconds to minutes, while the earthquakes themselves arrive as a sparse, self-exciting occurrence process, with inter-event intervals from seconds to months. Respecting both scales in one training objective is the open challenge. Foundation models (FMs) for seismic waveforms instead import their learning targets from audio and vision, where every target is computed from a single fixed-length window. Such a window holds one realization of the wave propagation (an earthquake trace), never the occurrence process that decides when earthquakes happen. We develop nine model designs sharing one Mamba encoder-decoder, trained on a fixed corpus of 3M windows, varying the self-supervised objective, quantization bottleneck and mixer; the two key comparisons are architecture-matched, varying the objective alone. Each checkpoint is evaluated frozen on three downstream tasks. Varying the target redistributes downstream skill rather than raising it, and no checkpoint wins everywhere. Adding masking pressure to our strongest classifier re-specializes it rather than making it a generalist: it gains 6.9 points of few-shot P-wave picking precision and loses 9.6 points of few-shot classification accuracy. By the standard imported from language and vision, these models are feature extractors rather than FMs. Controls on capacity, architecture, and domain proximity do not explain why skill moves rather than rises. Against task-specific baselines the advantage is regime-dependent: our checkpoints beat them few-shot on classification and ground-motion prediction, and trail them on full data. We argue the binding constraint is the target, and that settling the question requires a learning target defined across events, the timescale that makes seismicity stochastic in the first place.
Oct 27, 2026

Timestral

We introduce Timestral, a family of four time series foundation models for probabilistic forecasting, with 5M, 28M, 123M, and 209M parameters. Timestral sets a new state of the art on GIFT-Eval, TIME, and fev-bench, where Timestral-209M and Timestral-123M take the top two places on every metric. We achieve this with three ingredients: Decoupled Horizon Forecasting (DHF), which separates prediction tokens from the causally encoded context, so that longer horizons are trained without masking context and each observed value is both a forecast target and context for later forecasts; a covariate framework that natively supports known-future covariates within a causal architecture; and a two-stage pipeline that pretrains exclusively on a diverse, large-scale synthetic corpus before post-training on a mix of real and synthetic data with extended context. This report describes the model design and its benchmark results. Weights for all four models are available on Hugging Face and inference code on GitHub.
Oct 27, 2026

TRIDENT-Seq: Profile-Matched Prefix-Future Contrast for Label-Efficient Payment Risk Modeling

Ant International accumulates large volumes of transaction history across users and payment businesses. Confirmed fraud and chargeback outcomes, by contrast, are scarce, delayed, and specific to each business. A single encoder pretrained on unlabeled histories could serve all businesses, but only if pretraining captures the behavior that separates risk. Coarse activity differences can make contrastive pretraining on transaction logs too easy: transaction volume, amount, and timing can distinguish a history's continuation from arbitrary candidates without modeling how payment resources are used. We present TRIDENT-Seq, a transaction foundation model trained with Prefix-Future Contrast (PFC). PFC matches each prefix to a summary of its subsequent behavior among candidates with similar activity profiles; activity prediction and dispersion regularization act on the encoder output. Across three production payment businesses, the pretrained encoder improves average precision (AP) and recall at a 0.5% intervention budget over architecture-matched scratch training in all six business-budget settings. With a fixed validation set, it exceeds full-label scratch training in mean AP and recall using 10% of the training labels on Wallet and 25% on Shopping and SME. Pretraining only on another business's histories still helps. Replacing profile-matched candidates with random ones cuts recall from 24.62% to 5.64%. In Ant International's deployment evaluation, recorded amount-based chargeback and overall loss rates are 37.7% and 37.9% lower than those of the previous XGBoost-based system under the same intervention budget.
Oct 27, 2026

Learning from Points and Patches: A Pre-trained Hybrid Attention Transformer for Time Series Analysis

Time-series data generated by Web services and connected systems support applications such as user activity analysis, service monitoring, and resource management. However, learning transferable representations from heterogeneous time series remains challenging due to diverse temporal patterns, variable sequence lengths, and limited high-quality annotations. Despite the success of pre-training in natural language processing and computer vision, effective time-series pre-training requires strategies tailored to temporal data. We propose TimeHAT (Hybrid Attention Transformer), a general-purpose pre-training framework for time-series analysis that addresses two complementary questions: what to learn and how to learn. For what to learn, we introduce a Unified Point-Patch Masking Strategy (UPPMS) that integrates point-level and patch-level supervision within a unified framework for contextual reconstruction and causal extrapolation, enabling the model to capture both fine-grained temporal variations and broader temporal patterns. For how to learn, we develop a Hybrid Attention Encoder (HAE) that combines prefix-constrained bidirectional attention in initial layers with directional hybrid attention in deeper layers, aiming to preserve effective contextual interactions while mitigating rank degeneration. We evaluate TimeHAT across real-world datasets and four downstream tasks: classification, long-term forecasting, imputation, and anomaly detection. Experiments cover fine-tuning, few-shot, and zero-shot transfer settings, demonstrating competitive performance across diverse time-series analysis tasks. Code is available at https://github.com/TangPeiwang/TimeHAT.
Oct 26, 2026

Self-Supervised Pretraining for PHM: What Transfers, What Does Not, and What Federating It Costs

Self-supervised pretraining has produced capable time-series foundation models, but its value for Prognostics and Health Management (PHM) has not been measured under leakage-controlled conditions. We measure it on a corpus of 47 public PHM datasets, harmonized to a single tensor contract and partitioned so that 11 datasets are withheld from pretraining in full. Transfer turns out to be task-dependent: a compact channel-independent Transformer encoder (801,808 parameters) pretrained with masked patch prediction raises fault-classification macro-F1 from 0.350 to 0.812 on CWRU and from 0.569 to 0.951 on HSG18, yet on CMAPSS remaining-useful-life estimation no pretrained checkpoint reliably improves upon a from-scratch baseline. We pretrain under two regimes---centralized, and simulated cross-silo federated with ten clients grouped by physical domain---holding architecture, objective, source data and optimization budget fixed. Two further findings are negative. At matched budget the centralized encoder leads by 15.5--20.2 macro-F1 points in four of four comparisons, although the comparison is confounded by a learning-rate schedule that favours the centralized arm and rests on a single pretraining run per regime; on RUL the two regimes are indistinguishable. And the intrinsic SSL loss is not a valid selection criterion: the centralized encoder more than halves the federated reconstruction loss without a corresponding gap on RUL, and a server-momentum sweep moves downstream validation macro-F1 by up to 0.29 while the SSL loss varies by 0.021. Transfer also hinges on the backbone learning rate: the same checkpoint yields 0.169 or 0.829 macro-F1 on CWRU depending on a single order of magnitude. We release the corpus pipeline, the protocol and the code.
Oct 26, 2026

TimeWorld: Towards Action-Conditioned Multimodal World Models for Time Series Forecasting

Time series forecasting supports decision-making in web applications such as online retail and network management, where planned actions can reshape future dynamics. World models provide a natural paradigm for predicting such action-dependent changes beyond extrapolating historical patterns. However, existing time series world models (TSWMs) typically rely on fixed, domain-specific covariate schemas that limit flexible intervention specification, while multimodal forecasters treat text primarily as passive predictive context. To investigate these gaps, we introduce TSWM-Bench, a cross-domain benchmark for multimodal TSWMs spanning seven real-world datasets, with textual interventions grounded in recorded action--outcome data. Our evaluation reveals that conventional multimodal forecasters underperform history-only time series foundation models (TSFMs) despite dataset-specific training, whereas simple text-prefix adaptations improve forecasting across all evaluated TSFM backbones. These findings suggest that pretrained TSFMs encode implicit world knowledge of temporal dynamics, offering a strong foundation for action-conditioned modeling beyond conventional multimodal fusion. Building on this insight, we propose TimeWorld, a plug-in framework that adapts pretrained TSFMs through explicit action-response modeling. Specifically, the Prototype-Guided Latent Modulation (PLM) module grounds textual actions in temporal dynamics by composing representative response prototypes to modulate future representations based on the action and historical patterns. The Action Contrastive Learning (ACL) component strengthens action-specific response modeling by contrasting these updates under alternative interventions sharing the same history, distinguishing action types while preserving relative intensity relationships.
Oct 25, 2026

Dynamic Multi-Step Decoding for Time-Series Foundation Models

Time-series forecasting models typically use a fixed decoding strategy at inference time: either one-shot prediction of the full horizon or autoregressive rollout over atomic chunks. While one-shot decoding reduces error accumulation, it can discard useful sequential structure; conversely, autoregressive decoding preserves temporal dependencies but may compound errors. We argue that the optimal decoding strategy is not fixed across inputs, but should be selected adaptively for each forecasting instance. We propose a dynamic strategy selector that treats the forecast rollout schedule as an inference-time decision problem for frozen pretrained time-series foundation models. The selector is a learned network that ranks candidate rollout schedules using hidden states of the backbone, and is trained on historical windows of the evaluated tasks for each backbone. To make the candidate set tractable over the exponentially large space of horizon partitions, we use dynamic programming that reduces candidate evaluation to quadratic complexity while preserving much of the oracle gain. At deployment, five independently trained selectors vote on each decision, and the policy follows their consensus, falling back to the backbone's one-shot forecast where they disagree. The construction requires only the backbone's hidden states and lightweight selectors, with no fine-tuning of the forecasting backbone. On fev-bench with a frozen Chronos-2 backbone, agreement-gated selection improves point-forecast skill (MASE) over one-shot decoding with statistical significance, mostly on longer horizons. It recovers about 6-8% of the gain that the same dynamic program attains with ground-truth selection. With a frozen Moirai-1.1-R-small backbone, the improvement is also significant on the full benchmark, and for both backbones we claim gains in point accuracy only.
Oct 24, 2026

Unlocking the Specialization Capability of PPG Foundation Models for Wearable Signals through Structured Pruning and Adaptive Fine-Tuning

Photoplethysmography (PPG) plays a central role in wearable health monitoring and clinical decision support. However, existing approaches to universal PPG representation learning are predominantly trained on high-quality large-scale datasets obtained from oximeters in hospital settings and often struggle to adapt to sparse and low-quality real-world datasets, thereby limiting their generalization to clinically relevant tasks on prevalent PPG data such as collected from wearable devices. Surprisingly, even after full fine-tuning, PPG foundation models cannot consistently outperform general-purpose time-series models. Therefore, a key question is how to effectively adapt PPG foundation models to target prediction tasks. Through empirical studies on various PPG foundation models, we find that pretrained models often exhibit inherent sparsity and computational redundancy, suggesting that PPG foundation models have learned to activate task-relevant network substructures for different prediction tasks. To preserve this valuable prior knowledge, we propose a structured pruning-and-fine-tuning framework, SPAFT, which regularizes the adaptive fine-tuning process by constraining it to a more relevant and compact parameter space based on parameter importance and structural redundancy. Extensive experiments across four PPG foundation models, five benchmark datasets, and ten downstream tasks demonstrate that SPAFT achieves win rates of 82.8% and 87.5% on regression and classification tasks, respectively. This prune-then-finetune paradigm often enables PPG foundation models to achieve state-of-the-art performance, laying a promising foundation for the real-world clinical deployment of PPG foundation models and the development of personalized PPG foundation models.
Oct 23, 2026

Sustainability of TSFM Deployment and Pretraining: From Carbon Footprint Analysis to Mitigation via Pruning

Time Series Foundation Models (TSFMs) offer strong zero-shot forecasting capabilities, but their high pretraining cost raises a practical sustainability question: does this upfront investment pay off over a full deployment lifecycle, compared to supervised standalone models that must be periodically retrained? We model the cumulative carbon cost of different approaches across a realistic industrial scenario: Chronos-2 amortizes its pretraining cost within months for large or frequently queried deployments, but a classical model remains cheaper indefinitely for small or rarely queried ones. For cases where reducing the TSFM's own cost matters, we show that removing a single encoder block (7.9% of parameters) via a gradient-free pruning criterion yields a consistent 6 to 9% reduction in FLOPs, CO2 emissions, and latency, at essentially no cost to zero-shot forecasting performance across a 25-task benchmark.
Oct 22, 2026

Forecasting Associations, Not Interventions: A Preregistered Causal Audit of Covariate-Aware Time Series Foundation Models

Covariate-aware time series foundation models (TSFMs) accept a future path for a control variable and return a forecast, inviting what-if queries. The query is interventional; the models are trained on observational data. We determine whether their answers are the conditional slope ∂E[Y∣X]\partial E[Y|X] or the interventional effect ∂E[Y∣do(X)]\partial E[Y|do(X)], at three levels. Theory: a Bayes-optimal observational predictor answers interventional queries associationally unless the adjustment set is available and used; for in-context least squares we derive the resulting confounding gap -- a model's effect estimate under an observational minus under a randomized context -- in closed form, a parameter-free reference line. Behaviour: on a preregistered grid with exact counterfactual ground truth, Chronos-2, TiRex-2, TabPFN-TS and TimesFM-2.5 all track this line (R2R^2 0.90-1.00) and respond more confounded than regression itself (slopes 1.3-1.8), on linear and nonlinear structural models. Pairs of models with identical observational laws prove this is not an identifiability limit: when the effect is identifiable, least squares recovers 100% of the true effect difference, Chronos-2 75%, TiRex-2 34%. Representation: linear probes decode the data-generating regime (94-96%) and the true effect (R2R^2 0.7-0.9) from embeddings that the forecasts ignore. On two physical systems with randomized actuators, pretrained models under-respond by 2-5x and ignore in-window confounding. Minutes of interventional post-training on 3,000 synthetic series cut the gap by up to 83% and transfer outside the training prior. Its forecasting cost is a replay-distribution effect: replaying the model's own pretraining prior cuts it from +47% to +10% MASE, and with low-rank adaptation yields a recipe meeting the preregistered criterion over three seeds (52-61% gap reduction at +1-3% cost) while improving the DoTime benchmark.
Oct 22, 2026

Zero-Shot Forecasting of In-Situ Soil Moisture and Temperature with a Time-Series Foundation Model

Soil moisture and temperature govern plant-water availability and microbial activity, making them key indicators for evaluating land-restoration efforts on degraded landscapes. However, the physical and data-driven models normally used to forecast them both need information that a newly instrumented restoration site does not yet have: hydraulic parameters for the former and months of per-site training data for the latter. We evaluate TimesFM 1, a 500M-parameter time-series foundation model, zero-shot on 15-minute telemetry from 27 sensor stations across three land-management regimes (Control, Treatment, Scald) at a semi-arid restoration site, forecasting 5 cm and 30 cm soil moisture and enclosure temperature at seven lead times from 1 h to 168 h. Using gap-aware rolling-origin evaluation across sensors, we find that TimesFM beats the best rule-based statistical forecasting baselines (persistence, seasonal-naive, or 7-day climatology) by 21-76% at sub-daily lead times but loses by 6-79% beyond a day, a pattern explained by a diurnal-alignment mechanism that holds for both soil-moisture depths and temperature. A CatBoost baseline trained walk-forward on the same origins with full history and weather covariates never wins overall, and we show that its weakness is partly a cold-start effect specific to short deployments. We additionally quantify the inference cost of zero-shot forecasting on commodity CPU hardware. Our results highlight a practical trade-off between zero-shot forecasting capability, predictive performance, and computational efficiency.
Oct 22, 2026

Random Initialization as a Baseline for Time Series Foundation Models

Time series foundation models (TSFMs) use self-supervised pretraining to learn reusable representations for downstream tasks such as multivariate time series classification (MTSC). We examine how much pretrained weights contribute beyond the architecture itself by comparing seven publicly released instances from four TSFM families with architecture-matched randomly initialized counterparts on the UEA MTSC archive. For each instance, five random initializations are evaluated using four classifiers on frozen representations: ridge classification, logistic regression, RBF SVM, and random forest. After model-specific exclusions due to computational constraints, the study comprises 183 paired dataset-level comparisons. Pretraining improves mean accuracy for all seven instances, but the gains are modest, ranging from 1.43 to 5.19 percentage points, with an overall average improvement of 2.58 points. Three of seven per-instance Wilcoxon signed-rank tests have raw pp-values below 0.05, but only NuTime remains significant after Holm correction. A secondary pooled test across all 183 pairs gives p<0.001p < 0.001, while the corrected per-instance tests remain the primary evidence. We also compare with ROCKET, Hydra, and Quant, which remain competitive without requiring large-scale pretraining. These findings suggest that MTSC evaluations should distinguish the effect of pretrained weights from that of the underlying architecture and should include both architecture-matched random initialization and strong non-pretrained baselines.
Oct 21, 2026

Rethinking Agentic Model Orchestration for Time Series Forecasting

Time series foundation models (TSFMs) perform unevenly across forecasting tasks. Recent work uses agentic model orchestration (AMO) to select or combine TSFMs, often outperforming individual forecasters. Yet these gains may reflect stronger or more complementary candidates rather than better orchestration. To disentangle these effects, we introduce FairAMO, a controlled evaluation protocol with frozen candidate forecasts. Experiments on GIFT-Eval show that AMO improves accuracy, but its gains are bounded by the candidate TSFMs. The best orchestrators reduce MASE by less than 6% relative to the best single TSFM across all tasks. These gains depend on complementarity (different candidates excel on different tasks) rather than the number of candidates. Adding members yields diminishing returns, and a redundant member can erase these gains. The value of historical evidence also saturates quickly. A few recent windows capture all gains, and further history adds none. These findings separate system-level forecasting gains from the contribution of orchestration itself.
Oct 20, 2026

Understanding the Inner Workings of a State-Space Foundation Model for Limit Order Books

Autoregressive foundation models offer state-of-the-art capabilities to simulate the market by generating the message flow of limit order books (LOBs). Yet, without understanding their inner workings, research directions are hard to narrow and practical use in real-world trading environments remains uncertain. This study therefore asks what these models actually learn about market microstructure. We find that, similarly to games where sequence models build an internal representation of the board, an 8.2M-parameter Mamba-3 state-space model trained autoregressively on the tokenized message stream of eight large-cap US equities learns an internal representation of the market and its dynamics. In particular, order-flow imbalance, a central driver of short-horizon price moves, is decodable at R2≈0.59R^2 \approx 0.59, far above controls and baselines. This representation extends to unseen stocks, suggesting that these models learn general representations rather than ticker-specific patterns. The model also generalizes across metrics beyond order-flow imbalance: results hold across traded volume and volatility. Additionally, the representation is causally steerable: pushing the state along a feature's axis shifts the model's next-message beliefs as that feature implies, monotonically in dose, consistent with the emergence of a world model.
Oct 20, 2026

Turning TS2Vec into a Foundation Model for Time Series Classification

TS2Vec is one of the strongest self-supervised methods for time series classification, introduced in 2022. It is designed to be trained anew on every dataset: self-supervised pre-training followed by linear probing. Since then, the field has shifted towards foundation models that reuse one encoder across datasets. Yet TS2Vec offers a compact convolutional architecture that operates directly on temporal samples and accepts variable-length inputs without resampling. We ask whether these properties can be retained when the encoder is pre-trained only once, and show that they can: TS2Vec can essentially be turned into a foundation model. Three changes suffice: pre-train on a synthetic corpus, feed the encoder a raw amplitude view and an instance-normalized shape view of the input through two small stems, and distill one specialist encoder per view into a single encoder. The result, TS2Vec-FM, is a 1.6M-parameter dilated convolutional network that, kept frozen, reaches 83.3% on UCR-128 and 75.3% on UEA-27 under linear probing, within 0.3 pp of MantisV2, the current state of the art, and above TS2Vec trained separately on each target dataset. What we find most interesting is how different TS2Vec-FM is from other foundation models in design: near-parity is reached without attention, without a patch tokenizer, and without intermediate-layer extraction, using an "old-school" convolutional encoder and a two-layer MLP stem.
Oct 20, 2026

Parameter-Efficient Adaptation of Time Series Foundation Models for Electrical Load Forecasting

Accurate electrical load forecasting supports power-system operation across countries, regions, cities, and buildings, but repeated development of target-specific models is costly. Time series foundation models (TSFMs) provide reusable pretrained backbones, although multi-domain pretraining may not fully capture load-specific characteristics. This paper compares pretrained inference, full fine-tuning, and low-rank adaptation (LoRA) in terms of forecasting accuracy and adaptation efficiency across four electrical load aggregation levels. MOIRAI 1.0 Large serves as the primary backbone, while Chronos-2 assesses whether the findings transfer across model families. The evaluation covers both in-distribution (ID) and out-of-distribution (OOD) settings and examines whether compact adaptation provides consistent gains over strong pretrained checkpoints. Pretrained foundation models outperform the implemented seasonal naive and XGBoost baselines across the evaluated aggregation levels. On MOIRAI Large, rank-16 LoRA reduces normalized mean absolute error (NMAE) by 7.5% on average while updating 3.146 million parameters, about 1% of the model, and retains most of the improvement from full fine-tuning. Chronos-2 achieves lower pretrained NMAE than MOIRAI, while its relative adaptation gains remain smaller overall. OOD results show that adaptation benefits vary by target, with the clearest gains for the local-area series and several buildings. Overall, LoRA provides a favorable accuracy-efficiency trade-off when adaptation yields a meaningful gain over a strong pretrained checkpoint.
Oct 20, 2026

Foundation-Style Pretraining for Scalable and Transferable Distributed PV Forecasting

The rapid growth of distributed photovoltaic (PV) generation is increasing the demand for accurate forecasting across large numbers of geographically dispersed stations. Existing approaches are commonly developed for individual sites or specific forecasting horizons, requiring repeated training and maintenance as new stations are integrated and limiting direct reuse at previously unseen sites. To address this scalability gap, this paper presents a domain-specific foundation-style pretraining framework that learns a shared forecasting model from heterogeneous multi-site and multi-horizon PV data. NWP-assisted forecasting is formulated as a sequence-completion problem that unifies historical PV power, meteorological observations, future numerical weather prediction (NWP) forecasts, and time encodings. Three pretraining objectives are designed and systematically evaluated, including forecast-aligned reconstruction, generic reconstruction, and mixed reconstruction. The mixed objective combines the operational structure of PV forecasting with random sequence reconstruction and provides the most balanced performance across forecasting accuracy, degraded NWP conditions, and downstream adaptation. Experiments on 479 distributed PV stations use 340 stations for model development and 139 spatially held-out stations for zero-shot evaluation. Across horizons from 6 hours to 7 days, the multi-site pretrained MR model outperforms representative station-specific and shared baselines, as well as generalist time-series models, including a recent large-scale pretrained foundation model. The pretrained model can also be directly deployed at new PV stations and subsequently adapted to region-specific operational objectives when local data become available, providing a scalable forecasting paradigm for large distributed PV fleets.
Oct 19, 2026

WhatIf-Bench: Benchmarking Intervention-Aware Time Series Forecasting

Recent time series foundation models (TSFMs) can consume covariates in-context, extending zero-shot forecasting to problems driven by external factors. This enables scenario analysis: forecasting how a target responds to events that set covariates to chosen future values. Whether they can do this reliably is an open question. We introduce WhatIf-Bench, a benchmark measuring how well forecasters use covariates to forecast a target's response to future events. WhatIf-Bench is semi-synthetic: time series come from mechanistic simulators of data centers, power grids, road traffic, and water networks, with controlled covariate-target relationships grounded in each system. Each task presents a history of events intervening on covariates, then asks to forecast a future event that either replays seen values (Replay) or sets them to unseen ones (Novel). Varying the number of covariates, historical events, and event length yields 37,665 tasks. Evaluating TSFMs, frontier LLMs, and simple regression baselines, we find the LLMs reach the highest skill (improvement over a naive baseline), ahead of the best TSFMs. Every model forecasts well on Replay futures (48 to 85% skill) but drops to -7 to 50% on Novel futures, with the regression baselines falling the most. Skill on Novel futures also falls as events set more covariates jointly. More historical events raise skill on Novel futures across the board, but large differences remain. WhatIf-Bench reveals generalization to Novel futures as the core challenge, and offers a testbed toward arbitrary scenario analysis.
Oct 19, 2026

Beyond Static Leaderboards: Capability-Focused Evaluation with Structural Diversity for Time Series Foundation Models

Static public benchmarks for time series foundation models (TSFMs) typically have fixed and narrow test distributions. Their repeated reuse encourages benchmark-specific tuning, making reported performance reflect adaptation to the fixed test set rather than genuine forecasting generalization. In addition, a single aggregate score cannot fully characterize model capabilities across distinct temporal structures and input relationships. We propose CaFE (Capability-Focused Evaluation), a refreshable evaluation extension grounded in original benchmark data that enriches static benchmarks with structural diversity and enriches single-score black-box evaluation with multi-dimensional, capability-oriented assessment. CaFE first profiles the original benchmark and defines eight Structural Features (SFs) that characterize its temporal structure. For each SF, it resamples the original data according to its structural support, removes edge-of-distribution samples, and augments the retained samples by varying the SF's salience, producing samples with different structural profiles. After filtering out out-of-distribution samples while preserving diversity, CaFE yields a refreshed evaluation set. TSFMs are then evaluated on this set using both conventional forecast-error metrics and intervention-response metrics, giving each model a per-feature capability profile alongside endpoint accuracy. Experiments on GIFT-Eval and FEV-Mini20 show that CaFE yields stable evaluation results and reveals per-feature strengths and weaknesses that a single aggregate score hides. Moreover, on genuine data filtered by structural condition, models selected by CaFE outperform those chosen by the aggregate leaderboard. CaFE therefore makes model selection conditional on the structural response of interest.
Oct 19, 2026

From Retrieval to Recognition:Knowledge-Enhanced Post-training of Time Series Foundation Models for Classification

Time series foundation models (TSFMs) have achieved notable progress in classification tasks through large-scale pretraining. However, their post-training practice mainly focuses on implicit parameter fine-tuning, where task-relevant information is entangled with abundant irrelevant knowledge in vast parameter spaces. This may hinder TSFMs from precisely activating useful knowledge for downstream classification. To this end, we propose Retrieval-to-Recognition (R2R), a plug-and-play post-training paradigm that adapts pretrained TSFMs to time series classification. Rather than pretraining a new foundation model, R2R improves how existing TSFMs are post-trained: it introduces a decoupled query encoder to extract discriminative patterns and retrieve instance-level numerical knowledge from a constructed knowledge base. The retrieved explicit knowledge is then fused with the TSFM's implicit parametric embeddings, enabling joint knowledge transfer at both the parameter and feature levels. In this way, R2R allows TSFMs to recall task-relevant patterns learned during pretraining, improving the efficiency of knowledge utilization in downstream classification transfer. Experiments on 158 time series datasets show that R2R consistently improves the classification performance of state-of-the-art TSFMs. Ablation and knowledge-enhanced studies further validate the effectiveness of the proposed framework.
Oct 19, 2026

Adversarial Post-Training of Foundation Models for High-Frequency Trading

While autoregressive foundation models trained on limit order book (LOB) data show state-of-the-art performance for market simulation, pre-training with cross-entropy neither penalizes the compounding of sampling errors over free-running rollouts nor targets desired market properties. We introduce a post-training method for LOB foundation models using a Wasserstein generative adversarial network (GAN) objective. A critic, trained online on rollouts through the true, non-differentiable matching engine, guides updates from evolution strategies over rank-4 LoRA perturbations of the pre-trained generator, under a KL trust region. For efficiency, we use EGGROLL (Evolution Guided GeneRal Optimisation via Low-rank Learning), a memory-efficient variant of evolution strategies. On a held-out panel, adversarial post-training improves LOB-Bench realism by 17.6% (WS-21) and 17.0% (Mean-L1L_1), while selection-matched random-perturbation and compute-matched continued-pre-training controls yield no improvement. Post-training is thus a necessary stage that more pre-training does not replace. Per-position divergence measurements show that it slows error accumulation over 500-message rollouts from 3.1×3.1\times to 2.3×2.3\times. Adversarial post-training is also inexpensive, requiring under 50M training tokens and 12 GPU-hours per seed, three orders of magnitude below the 48B-token pre-training budget. Ablations over critic inputs, optimizers, and the population-rollout trade-off, together with an analysis of reward hacking, yield practical design guidelines for adversarial post-training.
Oct 18, 2026

Forecasting Retail Sales at Every Grain: Trees, Foundation Models, Transformers and World Models Compared on Four Public Data Sets

A retailer makes decisions from the same demand at different grains: store orders by item and day, promotions by item and week, category reviews and supplier terms. This paper asks whether one forecast, built at the finest grain and summed to every coarser grain, can serve them all, and which model does that best. It compares LightGBM, pretrained foundation models (TimesFM, Chronos-Bolt, Timestral), the Temporal Fusion Transformer and a world model that carries each item's state forward under the price and promotion plan, on M5, six Dominick's categories, Breakfast at the Frat and Corporacion Favorita, with the world model and LightGBM tuned by a common search. When to use the average: the average of the tuned world model and LightGBM, summed to every grain, has the lowest error in 22 of 27 grain and data set cases and is never the least accurate on the final held-out periods. When to use the world model: on existing items kept out of training it loses at most 0.040 in scaled error, against up to 0.311 for LightGBM, and on lines with short history it is clearly more accurate on six of nine data sets. When to use LightGBM: on heavily promoted lines, fast sellers and lines with long history (six of nine data sets each) and in price-cut weeks. Inconclusive: lines never promoted and slow lines are close to even, and 52 weeks ahead the average is the most accurate on five of seven data sets, with the world model ahead of LightGBM on two of seven. The findings hold for these data sets and the price and promotion levers; anyone adopting them should rerun every comparison on their own data.
Oct 16, 2026

Morpheus-2.0: Multiscale Probabilistic Forecasting with Dual-Axis Attention

A universal forecaster is trained once and applied everywhere, so it must serve tasks whose schemas, forecast horizons, and temporal resolutions all differ. The number of targets and covariates changes from dataset to dataset, horizons vary by orders of magnitude, and sampling rates range from seconds to years. Existing time-series foundation models address these only in part, since they isolate each series or visit the two grid relations in separate blocks, and they fix the horizon and resolution before training begins. We introduce Morpheus-2.0, a compact encoder-only foundation model that represents targets and covariates on a single variable-time grid. Its dual-axis attention runs both relations side by side in every layer, one across variables at a timestep and one across time within a variable; a learned gate weights the two per layer, and an axis-aware rotary encoding preserves both grid coordinates. The forecast length is sampled during training, so one set of weights spans short to long horizons at no additional inference cost. Morpheus-2.0 Multiscale extends the model to several temporal resolutions at once, weighting them according to the series being forecast. Across GIFT-Eval, which spans several domains and frequencies, the two models give complementary operating points: Morpheus-2.0 attains 0.685 relative MASE with 81M parameters, while the larger Morpheus-2.0 Multiscale attains 0.646 relative MASE and 0.455 relative CRPS, the lowest of any published system on both metrics.
Oct 16, 2026

Perceive, Solve, Reason: Progressive Capability Formation in Time Series-LLMs

Large language models (LLMs) can now perceive, describe, and reason about images and audio, yet time series understanding remains underestimated: most work uses LLMs as forecasters and classifiers, reasoning studies are few and narrow, and no suite trains and evaluates the full range of capabilities. In this paper, we build the first such suite, with 337K training and 44K evaluation samples across three progressive capabilities, Perceive, Solve, and Reason, and use it to ask how time series should enter an LLM that must master all three, and in what order. We find that a dedicated encoder is essential: with one, a 4B model surpasses GPT-4o on every perception attribute we test, whereas textualizing the series costs 8--10 points in reasoning accuracy. The capabilities then build on one another: learning them in order beats mixing the same data and rivals reasoning-only training with over three times the reasoning data, but only when an encoder is used. Analyses further reveal that capabilities are transferred through distinct network depths rather than simply stacked, and that a stage converged on its own benchmark is not necessarily ready for the next. Distilled into a simple recipe, these findings yield 4B TS-LLMs that, without any target-domain data, outperform four recent 7B TS-LLMs on seven of eight out-of-domain metrics, by up to 11 points. We therefore advocate treating time series as a first-class modality for LLMs, read through dedicated encoders and taught to perceive before they reason.