Time Series Analysis AI tools

AI Time Series Analysis platforms process and analyze temporal data patterns using artificial intelligence. These tools provide capabilities for trend analysis, forecasting, and anomaly detection in time-based data.

22 verified AI-first sites in Time Series Analysis.

Nixtla

Time series foundation model company behind TimeGPT and the open-source Nixtlaverse (StatsForecast, MLForecast, NeuralForecast). TimeGPT-2.1 and Nixtla Enterprise add multivariate and agentic forecasting plus anomaly detection; raised a $16M Series A in February 2026.

Pricing: Open-source libraries free; TimeGPT API free trial with usage-based and enterprise plans.

Tangent

Automated time series forecasting and anomaly detection engine that builds and rebuilds models on the fly without manual tuning. Available via API and as a Databricks runtime for high-throughput enterprise workloads.

Pricing: Custom enterprise pricing; contact Tangent.

GluonTS

Amazon open-source Python library for probabilistic time series forecasting with deep learning. Ships DeepAR, Transformer, and other models with training, backtesting, and evaluation tooling for univariate and multivariate series.

Pricing: Open source: Free.

NeuralProphet

Open-source PyTorch forecasting framework from Stanford that extends Prophet with neural components. Models trend, seasonality, holidays, lagged values, and external regressors with interpretable outputs; development has been quiet since early 2025.

Pricing: Open source: Free.

Anodot

AI business monitoring platform that learns normal behavior across time-series metrics and flags anomalies in revenue, payments, and operations in real time. Acquired by Glassbox in November 2025; still sold as a standalone product.

Pricing: Enterprise pricing with platform plans and integrations; contact sales.

TSFM

Unified API for time series foundation models including Chronos, TimesFM, and Moirai. Zero-shot forecasting, prediction intervals, anomaly detection, and imputation without provisioning GPUs or managing model weights.

Pricing: Free tier; see site for API pricing.

TimeCopilot

Open-source GenAI forecasting agent that combines LLMs with 30+ time-series foundation models (Chronos, Moirai, TimesFM, TimeGPT, and more). Natural-language forecasts, cross-validation, and anomaly detection through one API or CLI.

Pricing: Free and open source; LLM and model API keys as needed.

Granite FlowState

IBM Granite FlowState time series foundation model extending TinyTimeMixer-style efficiency to richer workloads. Compact pretrained checkpoint for zero-shot and few-shot forecasting on Hugging Face.

Pricing: Open weights on Hugging Face; see IBM Granite license.

Chronos

Amazon time series foundation model for zero-shot forecasting. Chronos-2 handles univariate, multivariate, and covariate-informed tasks with strong benchmark results and SageMaker deployment options.

Pricing: Open weights on Hugging Face; AWS SageMaker and AutoGluon Cloud deployment.

TimesFM

Google Research time series foundation model, available natively in BigQuery ML for zero-shot AI.FORECAST and anomaly detection in SQL. TimesFM-3 (August 2026) adds native multivariate forecasting and covariates, with open weights on Hugging Face.

Pricing: Included with BigQuery usage; open weights also on Hugging Face.

Moirai

Salesforce universal time series foundation model for zero-shot forecasting across domains, frequencies, and variable counts. Moirai 2.0 improves accuracy and speed on GIFT-Eval versus earlier releases.

Pricing: Open weights on Hugging Face; Uni2TS tooling for inference.

Toto

Datadog time series foundation model family for multivariate forecasting, tuned on observability metrics. Toto 2.0 scales from 4M to 2.5B parameters with strong GIFT-Eval and BOOM results.

Pricing: Open Apache 2.0 weights on Hugging Face.

NXAI

European lab behind TiRex time-series models built on xLSTM memory architectures, including TiRex-2. Targets industrial forecasting and edge or embedded deployment beyond transformer baselines.

Pricing: Research and partnership programs; contact NXAI.

MOMENT

Open time series foundation model family from CMU AutonLab for forecasting, classification, anomaly detection, and imputation. Pretrained on the Time Series Pile with weights on Hugging Face.

Pricing: Open weights on Hugging Face; Apache-licensed research code.

Timestral

Mistral's first family of time series foundation models for zero-shot forecasting, released in October 2026 in 5M, 28M, 123M, and 209M parameter sizes. It produces probabilistic multivariate forecasts across 21 quantiles, accepts past and known-future covariates, and handles up to 32,768 context steps on CPU or GPU. Mistral reported the 209M model first on GIFT-Eval, TIME, and fev-bench at release, with weights and code published under a modified MIT license.

Pricing: Free open weights (modified MIT license) on Hugging Face and GitHub.

The Forecasting Company

Research lab building the t0 family of forecasting foundation models, which read raw history plus related series and known-future inputs and return calibrated probabilistic forecasts. Its open-weights t0-gamma, released in October 2026, is reported as first on fev-bench, GIFT-Eval, and TIME at release, and runs locally through the open-source tfc-t0 library or through a hosted API. Requires no feature engineering or fine-tuning, and the Retrocast tool backtests a model against every series in a dataset.

Pricing: Open weights on Hugging Face; hosted API access via the company site.

TimeH2O

H2O.ai time series foundation model for zero-shot forecasting, released in October 2026 as TimeH2O-v1 with 108 million parameters. It forecasts univariate series and series with covariates such as price, promotions, or weather. H2O.ai reports first place by win rate on fev-bench's 100 tasks, ahead of Timestral-209M, t0-gamma, TimesFM-3, and Chronos-2; weights run locally through the h2ofm package.

Pricing: Non-commercial weights on Hugging Face; commercial use via H2O.ai.

Falcon TST

Ant International time series foundation model for zero-shot forecasting, with FalconTST 2.0 trained under the ORBIT regime and a Mixture-of-Experts architecture. The product site reports 2.5 billion parameters, 300 billion pretraining time points, and first-place zero-shot benchmark rank, with production use for FX and cashflow forecasting. Barclays, Citi, Deutsche Bank, and Standard Chartered have integrated the 2.0 model; an API trial ships through the falcon-tst package.

Pricing: API trial via the Falcon TST site and falcon-tst package; contact Ant International for commercial use.

YingLong

Alibaba research time series foundation model family for zero-shot forecasting, with open checkpoints from 6 million to 300 million parameters pretrained on 78 billion time points. YingLong is a bidirectional encoder trained with masked token recovery and delayed chain-of-thought output scaling, described in the 2025 Output Scaling paper. Weights are on Hugging Face under a CC BY 4.0 license and run through Hugging Face Transformers.

Pricing: Open weights on Hugging Face (CC BY 4.0).

Simons

Family of open time series foundation models from Abel AI Lab for zero-shot probabilistic forecasting. Four sizes from 25M to 297M parameters return nine forecast quantiles, take up to 32,768 steps of history, and accept multivariate series with known-future covariates. Weights ship under OpenMDW-1.1 with Apache-2.0 training code, synthetic data generators, and the public pretraining data, and the lab reports it ranks first on the M6 forecasting benchmark.

Pricing: Free open weights and code; no hosted API.

Flashmind

UAE lab building PDM, a family of zero-shot time-series forecasting foundation models in Nano, Pico, and Femto sizes. The models handle past and known-future covariates, missing observations, up to 16,336 past steps of context, and a 960-step native forecast horizon. PDM-Nano ranked first among zero-shot foundation models by normalized MASE on GIFT-Eval and TIME at launch, with submissions reproduced by the benchmark maintainers. Available through a console, API, and MCP server.

Pricing: Free to try; API access starts with a free key.

Continuous Intelligence

UK company building time series foundation models, with ci-proto as its debut family of five zero-shot forecasting models up to 167M parameters. Its attention-free time mixer scales linearly with history, letting the model read context windows reaching over two million observations. The largest model, ci-proto XL, took first place on the fev-bench leaderboard with a 51.5% skill score. The models return nine quantiles per target, support known-future inputs, run locally on CPU or GPU through a Python package, and are cleared for commercial use.

Pricing: Models available to use now; see the site for licensing and access.