Sold by: The Forecasting Company
Deployed on AWS
t0-alpha is a probabilistic time series forecasting foundation model. Give it a series' recent history and it returns multi-step quantile forecasts, with no per-series training or tuning. It works at any regular frequency from hourly to weekly, handles missing values, and can condition on known-future covariates such as holidays or planned promotions.
Overview
t0-alpha is a pre-trained foundation model for probabilistic time series forecasting. You pass a series' recent history and the number of steps you want, and it returns quantile forecasts for every step of the horizon, with no per-series training, fitting or hyperparameter tuning. One model serves every series you have. It works at any regular frequency from hourly to weekly, treats gaps as missing rather than as zeros, and can condition on known-future covariates such as holiday calendars, planned promotions or scheduled closures.
Because it returns a distribution rather than a single number, t0-alpha supports the decisions that actually depend on uncertainty: safety stock, reserve margin, staffing floors and capacity buffers can be set from the quantiles you ask for instead of from a guess about the error around a point forecast. It is built for teams running many series at once, in demand and inventory planning, workforce and capacity planning, energy load, logistics, retail sales and industrial sensor data, where building and retraining a model per series is not practical.
t0-alpha is pre-trained on publicly available time series datasets and on proprietary synthetic data generated in-house, and is used as-is. On GIFT-Eval, a public benchmark spanning 97 dataset configurations across multiple domains, frequencies and horizons, it achieves a mean weighted quantile loss (CRPS) of 0.491 and a MASE of 0.720 at the median. Deploy it as a real-time SageMaker endpoint for interactive and on-demand forecasting, at roughly 71 ms per request on a single ml.g4dn.xlarge, or as a batch transform job to score many series on a schedule. The container runs network-isolated inside your own AWS account, so the series you forecast never leave it.
Highlights
- No per-series training: T0 is a pre-trained time series forecasting foundation model, so each series is forecast from its own recent history alone. There is no model to fit, tune or retrain as series are added or churn.
- Quantile forecasts, not point estimates: every step of the horizon returns the quantile levels you ask for, so safety stock, reserve margins and staffing floors can be set from the distribution instead of guessed from a single line.
- Your data stays in your AWS account: the container runs network-isolated with the model weights bundled, as a real-time SageMaker endpoint at around 71 ms per request or as a batch transform job for scheduled scoring of many series.
Details
Sold by
Categories
Delivery method
Latest version
Deployed on AWS
New
Introducing multi-product solutions
You can now purchase comprehensive solutions tailored to use cases and industries.
Features and programs
Financing for AWS Marketplace purchases
AWS Marketplace now accepts line of credit payments through the PNC Vendor Finance program. This program is available to select AWS customers in the US, excluding NV, NC, ND, TN, & VT.
Pricing
Pricing is based on actual usage, with charges varying according to how much you consume. Subscriptions have no end date and may be canceled any time.
Additional AWS infrastructure costs may apply. Use the AWS Pricing Calculator to estimate your infrastructure costs.
Dimension | Description | Cost |
|---|---|---|
ml.g4dn.xlarge Inference (Batch) Recommended | Model inference on the ml.g4dn.xlarge instance type, batch mode | $100.00/host/hour |
inference.count.m.i.c Inference Pricing | inference.count.m.i.c Inference Pricing | $5.00/request |
ml.g4dn.2xlarge Inference (Batch) | Model inference on the ml.g4dn.2xlarge instance type, batch mode | $100.00/host/hour |
ml.g5.2xlarge Inference (Batch) | Model inference on the ml.g5.2xlarge instance type, batch mode | $100.00/host/hour |
ml.g5.4xlarge Inference (Batch) | Model inference on the ml.g5.4xlarge instance type, batch mode | $100.00/host/hour |
ml.g5.xlarge Inference (Batch) | Model inference on the ml.g5.xlarge instance type, batch mode | $100.00/host/hour |
Dimensions summary
This listing bills by usage, with two kinds of charges. The first is request-based inference pricing, where you pay per forecast request processed. The remaining five dimensions cover batch inference run on different instance sizes, billed per host hour. These instance options span two hardware families and range from smaller to larger configurations. You pick the instance that matches your workload, and cost scales with how many hours you run it. The request-based option and the hourly batch options are independent ways to consume the model, not tiers of one plan.
Top-of-mind questions for buyers
A request is one forecast call you send to the model's API. You describe your data once, then the model returns a forecast distribution for that input. Each call you submit counts toward the request-based charge. This option suits on-demand forecasting where you pay only for the calls you make.
The batch inference dimensions bill per host hour on the instance you run. Charges accrue while the instance is running, not by how many forecasts it processes. If you stop the instance, software host-hour charges stop. Underlying AWS resource fees may still apply separately depending on your setup.
The two models are independent ways to run the model, not stacked tiers. If you use the request-based API, you pay per forecast call. If you run batch inference on an instance, you pay per host hour. You can use either path, and each meters separately on your invoice.
theforecastingcompany.com
Helpful?
Vendor refund policy
No refunds.
Legal
Vendor terms and conditions
Upon subscribing to this product, you must acknowledge and agree to the terms and conditions outlined in the vendor's End User License Agreement (EULA) .
Content disclaimer
Vendors are responsible for their product descriptions and other product content. AWS does not warrant that vendors' product descriptions or other product content are accurate, complete, reliable, current, or error-free.
Delivery details
Amazon SageMaker model
An Amazon SageMaker model package is a pre-trained machine learning model ready to use without additional training. Use the model package to create a model on Amazon SageMaker for real-time inference or batch processing. Amazon SageMaker is a fully managed platform for building, training, and deploying machine learning models at scale.
Deploy the model on Amazon SageMaker AI using the following options:
Version release notes
Same model weights as version 0. The optional request field that selects quantile levels is now named "quantile_levels", matching the tfc-t0 SDK's predict(); the previous name "quantiles" is no longer accepted in requests. Responses are unchanged: "quantiles" holds the forecast and "quantile_levels" echoes the levels.
Additional details
Inputs
- Summary
JSON. A request describes one series:
- "context" is the list of past observations as floats, oldest first;
- "horizon" is the number of future steps to forecast;
- "quantile_levels" is an optional list of quantile levels, ascending and strictly between 0 and 1, defaulting to [0.1, 0.5, 0.9];
- "future_covariates" is an optional list of known-future signals shaped [n_covariates][len(context) + horizon].
Real-time endpoint
POST a single JSON request object, or a JSON list of request objects to forecast several series in one invocation. Content type application/json.
Batch transform
JSON Lines, one request object per line. Configure the job with split type Line, batch strategy SingleRecord and assemble with Line, so each line is sent to the model separately and the responses are reassembled in the same order.
Preprocessing
The payload carries values only, with no timestamps or frequency, so resample the series onto a regular grid before sending it and order it oldest first. Express a gap as NaN rather than as an omitted point or a zero. JSON has no NaN literal, so use an encoder that emits the non-standard NaN token: Python's standard-library json.dumps does this by default, as does the SageMaker SDK JSONSerializer. Strict encoders such as orjson refuse to emit NaN, so prefer the standard library for series that contain gaps.
- Limitations for input type
- Payload and runtime: up to 25 MB and 60 seconds per real-time invocation; up to 100 MB and 60 minutes per batch transform invocation. Instances: GPU only, for both real-time endpoints and batch transform jobs.
- Input MIME type
- application/json
https://github.com/theforecastingcompany/marketplace-examples/blob/main/samples/realtime-request.json
https://github.com/theforecastingcompany/marketplace-examples/blob/main/samples/batch-input.jsonl
Input data descriptions
The following table describes supported input data fields for real-time inference and batch transform.
Field name | Description | Constraints | Required |
|---|---|---|---|
context | Past observations of the series to forecast, as floating-point numbers ordered oldest first. The payload carries values only: no timestamps and no frequency are passed, so the model infers periodicity from the series itself. | Type: list of float. Length 1 to 16,384. Observations must be evenly spaced; resample irregular data onto a regular grid first. Send a gap as NaN, not null, since a literal null is rejected with HTTP 422. Truncate longer histories to the most recent window. | Yes |
horizon | Number of future steps to forecast. The response contains one set of quantiles for each step. | Type: integer. Range 1 to 2,048. Horizons beyond a single forward pass are produced by autoregressive rollout, so uncertainty widens with distance. | Yes |
quantile_levels | Quantile levels to return for every step of the horizon, named as in the tfc-t0 SDK's predict(). The response echoes the levels in the order requested. | Type: list of float. Default [0.1, 0.5, 0.9]. Each level strictly between 0 and 1, listed in ascending order. Levels the model was not trained on are interpolated. | No |
future_covariates | Known-future signals to condition the forecast on, such as holiday flags, planned promotions or scheduled closures. The meaning and encoding of the values are yours to define. | Type: list of list of float, shaped [n_covariates][len(context) + horizon]. Each covariate must cover the history and the whole horizon, so a signal known only historically cannot be used. Omit the field entirely rather than passing an empty list when no covariates apply. | No |
Support
Vendor support
E-mail us at support@theforecastingcompany.com , or use the dedicated support channels (e.g. Slack).
AWS infrastructure support
AWS Support is a one-on-one, fast-response support channel that is staffed 24x7x365 with experienced and technical support engineers. The service helps customers of all sizes and technical abilities to successfully utilize the products and features provided by Amazon Web Services.
Similar products
This data set include (seven 7) files that provides all information on ETFs listed in Canada.
Global Thin Film Solar Cell Market Size, Trends and Insights By Type (Cadmium Telluride, Amorphous Thin-film Silicon, Copper Indium Gallium Selenide, Microcrystalline Tandem Cells, Thin-film Polycrystalline Silicon, Others), By Installation (On-grid, Off-grid), By End-Use/Application (Residential, Commercial, Utility), and By Region - Global Industry Overview, Statistical Data, Competitive Analysis, Share, Outlook, and Forecast 2025 - 2034
Scalr is a cost-effective, drop-in replacement for Terraform Cloud with feature parity and better GitOps support. Full native CLI and API compatibility with TFC means you can migrate in minutes - your existing workflows, tools, and integrations just work. Purpose-built for Terraform and OpenTofu.
S&P Global Platts Commodity Risk Solutions provide risk valuation.
Customer reviews
No customer reviews yet
Be the first to review this product . We've partnered with PeerSpot to gather customer feedback. You can share your experience by writing or recording a review, or scheduling a call with a PeerSpot analyst.