Blog
Insights on time series foundation models, forecasting techniques, and the TSFM.ai platform.
TSFM vs LLM Research, Compared
We compared TSFM research against LLM research end-to-end — anchoring TSFM at Oct 2023 against LLM at Oct 2018, pulling publication-volume curves, model-release tempo, theme-by-theme alignment, and the published bibliometric record on both sides. The 5-year offset is real but the comparison is genuinely multi-axis: TSFMs are ahead on model releases and open weights, behind on publication velocity and scaling laws, and missing a ChatGPT-style consumer moment entirely.
2026
39 articlesHow TSFM Research Evolved: A Data-Driven Look at 2021–2026
679 papers, 43 release records, five years. We pulled the publication record from OpenAlex and assembled the model release record from primary sources to trace how the time-series foundation model field actually evolved — month by month, model by model.
ARFBench: Time-Series Models Have to Answer Questions Now
ARFBench turns observability telemetry into multiple-choice anomaly reasoning questions, exposing a new gap between forecasting accuracy and time-series question answering.
KairosHope: A Dual-Memory TSFM Built for Classification, Not Forecasting
KairosHope shifts the TSFM conversation from forecasting-first models to specialized classification, combining dual neural memory with classical time-series features, but public checkpoints are not available yet.
Kronos: A Domain-Specific Foundation Model for the Language of Financial Markets
Kronos is a decoder-only foundation model pre-trained on 12 billion K-line records from 45 global exchanges, with a specialized tokenizer that quantizes OHLCV bars into hierarchical coarse/fine subtokens. Accepted at AAAI 2026, the model reports 93% higher RankIC than the leading general-purpose TSFM on financial benchmarks and is open-sourced under MIT.
TSFM Scaling Laws in 2026: What Seven Families Tell Us About Bigger Time Series Models
For most of the time series foundation model era, bigger has not reliably meant better. By mid-2026 the empirical picture has changed: seven released TSFM families — from 1M-parameter Granite TTM up to 8.3B-parameter Timer-S1 — give the field enough scaling data points to talk seriously about scaling laws. Toto 2.0 and Xihe report monotonic scaling at five sizes each; TimesFM 2.5 went the other way and got smaller; Moirai held mostly at Base. This post walks through what each family proves, fits the data against the published TSFM scaling-law literature, and works out what 'compute-optimal' should actually mean when inference dominates total cost.
Xihe: A Scaling Family Built Around Hierarchical Interleaved Block Attention
Xihe is a five-checkpoint TSFM family (9.5M → 1.5B parameters) built around Hierarchical Interleaved Block Attention (HIBA), a sparse multi-scale attention layout that alternates intra-block and inter-block attention at hierarchical block sizes. The paper reports new top-of-leaderboard zero-shot numbers on GIFT-Eval and a smallest-model line that already beats many incumbents.
BVARs vs. TSFMs: Priors, Shocks, and the New Forecasting Stack
Bayesian VARs and time series foundation models solve different forecasting problems: BVARs encode structural priors and scenarios; TSFMs scale zero-shot forecasts across heterogeneous series.
Toto 2.0: Datadog Turns Observability Forecasting Into a Scaling Family
Datadog's Toto 2.0 release changes the Toto story from one 151M-parameter observability model into a scaling family. TSFM.ai hosts four launch checkpoints from 4M to 1B parameters, leaving 2.5B upstream-only for now, while the release adds u-muP scaling, alternating time/variate attention, and quantile forecasts.
Agricultural Commodity Forecasting: When Zero-Shot TSFMs Beat Futures-Based USDA Baselines
A new agricultural price forecasting study compares 17 methods on USDA ERS commodity prices from 1997-2025. The surprising result: zero-shot TSFMs take the top five monthly forecasting ranks, and Time-MoE beats futures-based USDA benchmarks on wheat and corn despite using only historical cash prices.
Aurora: What the First Multimodal Time Series Foundation Model Actually Does
Aurora (ICLR 2026) is the first pretrained TSFM to fuse text and visual structure at training time — and it benchmarks above Chronos, Moirai, and TimesFM on both deterministic and probabilistic metrics. Here's what the architecture actually does, what 'multimodal' means in practice, and how to run it today.
ICLR 2026 TSFM Roundup: 87 Time Series Papers and What They Mean for the Field
ICLR 2026 accepted 87 time series papers — the first year with a dedicated TSALM workshop. Here's what the research community spent the year building: multimodal pretraining, elucidated diffusion, prototype-guided generation, and a renewed focus on rigorous evaluation.
TEDM: What Happens When You Apply Elucidated Diffusion Models to Time Series
TEDM (ICLR 2026) adapts Karras et al.'s Elucidated Diffusion Models to time series by replacing hand-crafted noise schedules with data-derived covariance estimates — achieving linear inference complexity in forecast horizon and the fastest training time among diffusion-based forecasters.
CPU-Only Energy Load Forecasting: What a New ERCOT Benchmark Reveals About TSFMs Without GPUs
A new multi-dimensional benchmark runs four TSFMs on a consumer laptop—AMD Ryzen 7, 16 GB RAM, no GPU—against ERCOT hourly load data from 2020 to 2024. The headline finding: foundation models beat Prophet catastrophically at short context, cluster within statistical noise at 512-hour context, and diverge sharply on calibration. Here is what it means for practitioners who cannot assume GPU access.
iAmTime: Instruction-Conditioned In-Context Learning for Multi-Task Time Series Foundation Models
A new foundation model shows that time series tasks — forecasting, classification, anomaly detection, imputation — can all be specified at inference time through structured input-output demonstrations, without task-specific heads or fine-tuning. The model, iAmTime, achieves best-in-class CRPS on fev-bench and competitive performance on GIFT-Eval using only zero-shot, instruction-conditioned in-context learning.
TIME: Why Task-Centric Zero-Shot Evaluation Changes How We Compare TSFMs
Most TSFM benchmarks recycle legacy datasets, apply mechanical horizon rules, and report model rankings by coarse domain label. TIME (arXiv:2602.12147, February 2026) breaks all three habits. Built from 50 fresh datasets and 98 operationally-aligned forecasting tasks, it evaluates 12 TSFMs under strict zero-shot conditions and introduces a 7-feature binary pattern encoding that replaces static domain labels with intrinsic temporal structure. The result is a leaderboard where model rankings are stable at the top but shift substantially by stationarity and aggregation level — and where the finding that newer TSFM generations genuinely outperform their predecessors can be verified on contamination-free ground.
Cisco Time Series Model: Multiresolution Input as an Architecture Primitive
Cisco's CTSM 1.0 takes a different approach to long context: instead of extending the sequence length, it feeds the model two views of the same history at different resolutions — coarse hourly context for global patterns, fine minute-level context for local detail. On observability data the approach leads MASE by 16%, and on GIFT-Eval it lands third without test-set leakage.
IBM's March 2026 Time-Series Refresh: Four Models, Three Forecasters, and a Surprise
IBM dropped four time-series foundation models on March 31 — FlowState r1.1, TTM-r3, PatchTST-FM r1, and TSPulse r1. Between them they cover point forecasting, probabilistic forecasting, anomaly detection, classification, and imputation. Here is what changed architecturally and why it matters.
TFRBench: The First Benchmark for Evaluating Reasoning in Forecasting Systems
Nearly every time series benchmark measures whether a model predicts the right numbers. TFRBench (arXiv:2604.05364, April 2026, Google Cloud AI Research) asks a different question: can the model explain *why*? According to its authors, TFRBench is the first benchmark designed to evaluate the reasoning capabilities of forecasting systems — covering cross-channel dependencies, trends, seasonality, and external events — finding that off-the-shelf LLMs consistently struggle with temporal reasoning, while models prompted with high-quality reasoning traces improve from about 40.2% to 56.6% on the share of series beating the naive baseline. The implications touch deployment design decisions across any organization asking forecasting systems to justify their outputs.
QuitoBench: What a Billion-Scale Benchmark Reveals About When to Use Foundation Models
Most time series benchmarks organize datasets by domain label — traffic, energy, weather — which says little about why a series is hard to forecast. QuitoBench (arXiv:2603.26017, March 2026) reorganizes evaluation around TSF regimes: eight combinations of trend strength, seasonality strength, and forecastability. Benchmarking 10 models across 232,200 instances on the Quito corpus, it reports a context-length crossover, a forecastability dominance finding, and practical model-selection guidance that domain-based benchmarks tend to blur.
CoRA: A Plug-and-Play Adapter for Adding Cross-Channel Correlation to Any TSFM
Most time series foundation models treat each variate independently and miss the cross-channel dynamics that matter most in real multivariate workloads. CoRA — a Correlation-aware Adapter from East China Normal University and Huawei — is a lightweight plug-and-play module that adds three-type correlation modeling to any pretrained TSFM during fine-tuning, with linear inference complexity and consistent improvements across six base models on ten benchmark datasets.
Moirai-MoE: Token-Level Specialization for Time Series Foundation Models
Salesforce's Moirai-MoE replaces Moirai's frequency-specific projection layers with a sparse mixture of experts, enabling automatic token-level specialization. With only 11M activated parameters, Moirai-MoE-Small beats Moirai-Large across 29 Monash benchmarks, and Moirai-MoE-Base outperforms Chronos and TimesFM on zero-shot aggregate metrics.
TimesFM In-Context Fine-Tuning: Domain Adaptation Without Weight Updates
Google Research's In-Context Fine-Tuning (ICF) lets TimesFM adapt to new domains at inference time by prompting with related time series examples — no gradient updates, no training pipeline, roughly 7% better accuracy on aggregate across Monash benchmarks. Here's how it works and when to use it.
Causal Reasoning in TSFMs: From Forecasting to Decision Support
TSFMs tell you what will happen. Decision-makers need to know what would happen under different actions. Two new research directions — interventional priors and activation transplantation — are giving TSFMs their first taste of causal reasoning, opening a path from passive prediction to active decision support.
TSFMs vs World Models: Two Philosophies of Prediction
Yann LeCun's AMI Labs just raised $1B to build world models. DeepMind shipped Genie 3. Researchers are adapting JEPA for time series. Should practitioners care — or is this a different field solving a different problem? Here's a technical comparison of the two prediction paradigms.
Encoder-Only, Decoder-Only, and Encoder-Decoder: Architecture Tradeoffs in Time Series Foundation Models
Every major TSFM makes an architectural bet — encoder-only, decoder-only, or encoder-decoder. Each choice cascades into real differences in inference speed, horizon flexibility, and what the model can learn. Here's how to think about the tradeoffs.
IBM FlowState: The First Sampling-Rate-Invariant Time Series Foundation Model
IBM Research's FlowState abandons transformer attention entirely in favor of a State Space Model encoder and a Functional Basis Decoder that produces continuous-time forecasts. Under 10M parameters, it reaches the top of the GIFT-Eval zero-shot leaderboard — and can adapt to any sampling rate without retraining.
Impermanent: The First Live Benchmark for Temporal Generalization in TSFMs
Static benchmarks contaminate themselves over time. Impermanent is the first live benchmark that scores TSFM forecasts sequentially on a continuously updated data stream, testing whether foundation-model generalization actually holds once real time passes.
Timer-S1: Billion-Scale Time Series Forecasting with Serial Scaling
Timer-S1 from Tsinghua University introduces Serial Token Prediction and a sparse MoE architecture with 8.3B total parameters to achieve state-of-the-art forecasting on the GIFT-Eval benchmark.
Traditional Forecasting vs. TSFMs: The True Cost of Building and Maintaining Enterprise Forecast Pipelines
Enterprise forecasting pipelines take months to build, require specialized teams to maintain, and silently accumulate technical debt. Time series foundation models compress the entire lifecycle into an API call. Here's a honest comparison of both paths.
LoRA and PEFT for Time Series Foundation Models: A Technical Guide
A deep technical guide to applying Low-Rank Adaptation and parameter-efficient fine-tuning to time series foundation models — covering architecture-specific strategies, rank selection, layer targeting, catastrophic forgetting, and practical results across Chronos-2, MOMENT, Moirai, TimesFM, and more.
BOOM: Datadog's Observability Forecasting Benchmark
BOOM evaluates time series models on 2,807 real-world production monitoring series from Datadog. Here's how it works and why observability data demands its own benchmark.
Real-Time Streaming Inference with TSFMs: Moving from Batch to Continuous Forecasting
Most TSFM deployments run forecasts in hourly or daily batches, but a growing class of use cases demands continuous, low-latency predictions. Here's how to architect streaming inference pipelines for time series foundation models.
GIFT-Eval: Salesforce's Comprehensive TSFM Benchmark
GIFT-Eval tests foundation models across 23 datasets, 7 domains, and both univariate and multivariate settings. Here's what makes it one of the most thorough TSFM benchmarks available.
Giving Claude Time-Series Superpowers with MCP
Large language models can't forecast time series — but they don't have to. With MCP tool use, Claude can call specialized foundation models and return calibrated probabilistic forecasts in a single conversation turn.
The 2026 TSFM Toolkit: Which Foundation Model for Which Job?
With 18+ time series foundation models now available, choosing the right one for your workload is the real challenge. Here's a practitioner's decision framework.
FEV Bench: The Zero-Shot Forecasting Benchmark Explained
FEV Bench from AutoGluon evaluates foundation models on pure zero-shot forecasting across 29 diverse datasets. Here's how it works, what it measures, and why it matters.
Multimodal Time Series: Combining Numerical Data with Text and Natural Language Context
What if forecast models could read a description of your data alongside the numbers? Multimodal time series research is exploring how natural language context — domain descriptions, known events, business constraints — can improve forecasts, especially in zero-shot settings.
The TSFM Landscape in 2026: Trends and Predictions
As we enter 2026, the time series foundation model ecosystem is maturing rapidly. Here are the trends shaping the next year.
Network Traffic Forecasting and Telecom Capacity Planning with TSFMs
Telecom and network operators generate some of the richest time series data on earth — CDN throughput, cell tower load, DNS query rates — and time series foundation models can forecast it without per-metric training.
2025
21 articlesScaling TSFM Inference: GPU Optimization
Serving TSFMs at scale requires careful GPU optimization. Here's how we achieve sub-100ms latency for batch forecasting.
Toto: Datadog's Domain-Specific TSFM for Observability
Datadog's Toto is a 151M-parameter transformer trained on trillions of observability data points, purpose-built for forecasting infrastructure metrics like CPU utilization, error rates, and request latency.
TiRex: NX-AI's xLSTM Zero-Shot Forecasting Model
TiRex is NX-AI's 35M-parameter xLSTM forecasting model, built for strong zero-shot performance across short and long horizons with point and quantile outputs.
Smart Model Routing: Choosing the Best TSFM
Not every time series needs the same model. TSFM.ai's routing engine automatically selects the best foundation model for each request.
Predictive Maintenance in Manufacturing with TSFMs
Time series foundation models enable predictive maintenance across thousands of machine types without training individual models, catching bearing degradation, tool wear, and compressor decay before costly failures occur.
LLM Reprogramming for Time Series: Time-LLM's Prompt-as-Prefix Approach
Time-LLM reprograms a frozen LLaMA-7B to forecast time series by mapping patches into the LLM's embedding space and prepending natural language task descriptions, raising a provocative question about what large language models actually learn.
Supply Chain and Logistics Forecasting with TSFMs
Time series foundation models tackle the toughest supply chain forecasting problems, from the bullwhip effect to cold-start routes, enabling leaner inventory and fewer stockouts across complex logistics networks.
How Retailers Use TSFMs for Demand Planning
From inventory optimization to markdown pricing, time series foundation models are transforming how retailers forecast demand.
Healthcare Applications of Time Series Foundation Models
From ICU vital sign monitoring to hospital capacity planning, time series foundation models address healthcare's most pressing data challenges — starting with the chronic scarcity of labeled clinical data.
MOMENT: CMU's Model for Time Series Understanding
MOMENT from Carnegie Mellon is a multi-task time series foundation model that handles forecasting, classification, anomaly detection, and imputation.
The Founder's Crash Course on Time Series Foundation Models
You already ship forecasting in your product. Here's why time series foundation models change the economics, accuracy, and speed of everything you've built — and what to do about it.
Climate and Weather Forecasting with Time Series Foundation Models
Time series foundation models are finding a practical niche in climate and weather applications — not replacing physics-based models, but filling gaps they leave behind.
Context Length in TSFMs: How Much History Do Foundation Models Need?
Time series foundation models accept anywhere from 4K to 16K input time steps, but more context is not always better. We break down when longer history helps, when it hurts, and how to choose the right lookback for your data.
The State of Multivariate Forecasting in 2025
Multivariate time series forecasting remains one of the hardest problems in ML. Here's where foundation models stand in 2025.
Diffusion Models for Time Series: Inside Tsinghua's Sundial
Sundial applies flow-matching diffusion to time series forecasting, producing full predictive distributions from noise in fewer steps than standard diffusion models.
Covariates in Time Series Forecasting: A Practical Guide to Exogenous Regressors
Covariates like holidays, promotions, and weather can dramatically improve forecast accuracy. Here's how modern TSFMs handle them and when they actually help.
Fine-Tuning vs. Zero-Shot: When to Customize
Zero-shot TSFMs are powerful out of the box, but sometimes fine-tuning on your data delivers a meaningful accuracy boost. Here's how to decide.
Tiny Models, Big Results: IBM's Granite TTM and the MLP-Mixer Architecture for Time Series
IBM's Granite TTM packs competitive forecasting accuracy into roughly 1 million parameters by replacing attention with MLP-Mixer layers, enabling sub-100ms inference on CPU and opening the door to edge and serverless deployment.
Chronos v2: What's New and Why It Matters
Amazon's Chronos-Bolt brings major improvements: direct multi-step forecasting, much lower latency than the original Chronos-T5 stack, and stronger benchmark results.
Conformal Prediction: Calibrated Uncertainty Intervals for Time Series Foundation Models
Model-native prediction intervals are often miscalibrated. Conformal prediction provides a distribution-free wrapper that turns any forecaster's output into intervals with guaranteed coverage.
Case Study: Energy Demand Forecasting
How a European energy utility used TSFM.ai to improve demand forecasts by 23% compared to their existing gradient boosted tree pipeline.
2024
15 articlesIntroducing TSFM.ai: The Unified API for Time Series
We're launching TSFM.ai — a single API that gives you access to every major time series foundation model with automatic routing and optimization.
The Challenges of Benchmarking TSFMs
Benchmarking time series foundation models is harder than it looks. Here's why results often conflict and what the field is doing about it.
Time-MoE: Xiaohongshu's Mixture-of-Experts Architecture for Time Series Forecasting
Xiaohongshu's Time-MoE brings sparse mixture-of-experts to time series forecasting, activating only a subset of parameters per input to achieve large-model capacity at lower inference cost.
Lag-Llama: The Open-Source Time Series Foundation Model
Lag-Llama brings the decoder-only LLM architecture to time series with lag-based tokenization and distributional outputs.
Prediction Intervals vs. Point Forecasts
Why a single predicted number is rarely enough, and how prediction intervals help you make better decisions under uncertainty.
Synthetic Training Data for TSFMs: KernelSynth and Gaussian Process Augmentation
How KernelSynth uses Gaussian process priors with composed kernels to generate synthetic time series, and why roughly half of Chronos's training data is artificially generated.
Building Production Forecast Pipelines with TSFMs
A practical guide to moving time series foundation models from notebooks to production-grade forecasting systems.
Financial Time Series: Volatility Modeling and Risk Forecasting with TSFMs
Financial markets produce some of the most challenging time series data. Here's how time series foundation models handle volatility clustering, tail risk estimation, and regulatory risk forecasting.
Moirai: Salesforce's Universal Forecasting Transformer
Moirai from Salesforce introduces a universal forecasting transformer that handles variable frequencies, prediction lengths, and multivariate inputs.
Tokenization Strategies Compared: Quantization vs Patching vs Lag-Based
Time series foundation models must bridge the gap between continuous signals and discrete transformer inputs. We compare three dominant tokenization strategies — quantization, patching, and lag-based — and when each one works best.
Anomaly Detection with Time Series Foundation Models
Foundation models aren't just for forecasting — they're surprisingly effective at detecting anomalies in time series data.
TimesFM: Google's Approach to Time Series Foundation Models
Google's TimesFM family is a decoder-only foundation model line for time series, spanning the original 200M release and newer 2.0/2.5 checkpoints with longer context and optional quantile support.
Zero-Shot Forecasting: Why It Matters
Zero-shot forecasting lets you generate predictions on unseen time series without any training. Here's why that's a game-changer.
A Deep Dive into Amazon Chronos
How Amazon's Chronos turns time series forecasting into a language modeling problem using tokenized values and T5 architectures.
What Are Time Series Foundation Models?
An introduction to time series foundation models — what they are, how they work, and why they represent a paradigm shift in forecasting.