Timestral

Forecasting software

Forecasting software without the platform rollout.

Most forecasting tools are platforms: connectors, a warehouse model, a training pipeline, a governance layer. That is the right answer for a data team with a permanent forecasting program. It is the wrong answer when you have one CSV and a decision due this week.

This page compares the main paths — hosted CSV tools, open-source libraries, cloud platforms, open foundation models and spreadsheets — by the workflow they actually support. The goal is not to declare one winner. It is to help you choose the smallest tool that can produce a defensible range for your decision.

Try the CSV studioRead the API contract
ToolModelInputOutputSetupPriceBest for
Timestral StudioHosted Timestral-209M, zero-shotCSV upload or REST APIMedian, 80% band, 21 quantilesMinutes$9.90–$639 self-serveA fast forecast from one CSV, or an API without an ML stack
Nixtla / StatsForecastStatistical and ML models you choosePython data framesModel-dependent; intervals availableHours to daysOpen source; cloud is separateA Python team that wants control and scale
ProphetAdditive trend/seasonality modelPython or R data framesPoint forecast plus uncertainty intervalHoursOpen sourceAnalyst-driven series with clear trend and seasonality
Cloud forecasting platformsManaged AutoML and custom modelsWarehouse or cloud storageConfigurable forecasts and intervalsDays to weeksUsage-based; request a quoteLarge pipelines, governance, connectors and a platform team
Open foundation modelsChronos, TimesFM, Timestral and similarTensors or CSV through your own codeModel-dependentDaysModel weights are open; hosting is yoursML engineers who want to embed a model in their own product
SpreadsheetsManual formulas and chartsRows and cellsA number a person choseMinutesAlready paid forA one-off sanity check, not a repeatable forecast

Competitor capabilities and pricing change. The table describes the workflow each category is built for, not a live price list. Verify current terms with each vendor before committing.

The real decision: where does the forecast live?

Forecasting tools differ less on the model than on where the forecast lives. A spreadsheet lives in a person’s file. A library lives in a Python service. A cloud platform lives in a warehouse and a governance process. Timestral lives behind a CSV upload and a REST endpoint.

If the answer is for a person

Use a studio. Upload the CSV, read the chart, download the quantiles. No deployment, no notebook.

If the answer is for a product

Use an API. Authenticate with a key, send JSON, store the task id and the result file. Meter the calls.

If the answer is for a platform

Use a library or a cloud service. You need connectors, retraining, lineage and a team to own them.

When Timestral is the wrong choice

A good comparison includes the cases where the product should lose. Timestral is not the right tool if any of these is a hard requirement:

  • You need connectors to a warehouse, automatic retraining, lineage and approval workflows.
  • You need hierarchical reconciliation across thousands of SKUs, regions and channels.
  • You need causal inference or explicit promotion/price elasticities, not a statistical forecast.
  • You need a private deployment inside your own VPC or on-premises.
  • You need sub-100ms scoring at very high volume with a custom model.
  • You need a regulated data path that forbids sending the series to a third-party inference host.

When Timestral is a good fit

  • You can export the series to CSV and need a forecast today, not after a platform rollout.
  • You want a probabilistic output — a median and a range — rather than a single number.
  • You want the same forecast behind an API without building a model-serving stack.
  • You need a second opinion on an existing plan, with a transparent per-call price.
  • You need a zero-shot baseline before investing in custom training.

How to evaluate any forecasting tool

Do not choose on a demo. Hold out the last few periods of your own series and compare the forecast with what actually happened. Use three numbers:

MASE

Mean absolute error scaled by a naive baseline. Below 1 means the model beats the naive forecast; lower is better.

RMSE

Root mean squared error in your unit. Useful for absolute planning, but sensitive to large misses.

80% coverage

The share of actuals inside the low/high band. If it is far below 80%, the band is too narrow to plan with.

Also measure the boring parts: latency per call, cost per 1,000 calls, how failures are refunded, and what happens to your data. Those decide whether a tool survives production, not the demo chart. Our own benchmark and method are published on the benchmark page.

CSV forecast studio

Compare with your own CSV.

Run the same series through the hosted studio. The sample is airline passenger data; replace it with the series you would use in a vendor trial. The output is a median and an 80% band, with all 21 quantiles available for download.

1 · Paste or upload a CSV2 · Pick a horizon3 · Download the 21 quantiles
1 credit per API call · the web demo is free while the public Space is available

FAQ

What is the fastest way to forecast one CSV?+

Upload the CSV to Timestral Studio, choose a horizon, and run the forecast. There is no model training step and no Python environment. The output is a median and an 80% range; the API returns all 21 quantiles for production use.

Is Timestral a replacement for a forecasting platform?+

No. It is a focused CSV studio and API. It does not provide data connectors, hierarchical reconciliation, approval workflows, scheduled retraining or a private deployment out of the box. Those are reasons to choose a platform or build on an open library.

Should I use an open-source library instead?+

If you have a Python team, need full control over the model, and can own the infrastructure, an open library such as StatsForecast or Prophet is a strong choice. Timestral is for teams that want a working forecast and a documented API without building an ML stack.

How do I compare accuracy across tools?+

Backtest on your own series with a held-out window. Report MASE for scale-free error, RMSE for absolute error in your unit, and 80% coverage for the interval. A tool that wins on one metric but misses the interval is not automatically better for planning.

What does a forecast cost?+

Timestral uses credits: one credit per successful call. The cheapest self-serve tier is the Studio annual plan at $639 for 6,000 credits, which is about $106.50 per 1,000 calls. The entry one-time pack is $9.90 for 10 calls. Failed calls are refunded.

Can I use the API without a subscription?+

Yes. One-time credit packs never expire. Subscriptions add credits each billing period. Create a key in the account settings after signing in with an email code.

Start with one CSV.

No account needed for the studio. The API is metered; pricing starts at $9.90.

See pricingAPI reference