Project a metric forward, with an honest range around it
Pick a metric, such as weekly revenue or daily signups. clariBI tries several forecasting methods on your own history, keeps the one that would have predicted best, and shows the projection with ranges that say how sure it is.
Starter, Professional and Enterprise. Not included in the trial. The maths uses no language model.
How a forecast is made
Test on the past
clariBI hides the most recent part of your history, forecasts it from the rest, and compares the forecast with what really happened. It repeats this several times, moving forward each time. This is called a walk-forward backtest.
Keep the best method
About 18 methods take part, from "same as last week" to seasonal and trend models, a machine-learning model and an average of the best three. The one with the smallest error on your data is used.
Show the uncertainty
The errors from those tests set the ranges: the band that held 50, 80 and 95 percent of past outcomes. A wide band means the metric is hard to predict, which is worth knowing.
Which other metrics move with it
- clariBI compares the metric with the other metrics in your account, including with a delay of up to 14 days, so a metric that tends to move a week earlier shows up.
- Testing many pairs produces some matches by chance. clariBI corrects for this (false discovery rate held at 5 percent), so fewer coincidences are reported.
- Each driver is re-tested on resampled data. If the relationship does not hold up, it is marked as fragile.
- A driver moving with your metric does not prove it causes it. Treat the list as a place to look, not a conclusion.
Unusual points and changes of trend
- Anomalies: points far outside the normal range for that series, labelled by severity.
- Changepoints: dates where the trend changed slope, with the rate before and after.
- A short written summary of the result. It describes the numbers above; it does not produce them.
What you choose
| Setting | Options |
|---|---|
| What to forecast | A metric in a report, a widget on a dashboard, a goal's progress, or a time series from a connected source |
| Granularity and horizon | Hourly (up to 168 hours ahead), daily (90 days), weekly (52 weeks), monthly (24 months), quarterly (8 quarters) or yearly (5 years). 5 to 45 minute series are also supported |
| Schedule | Re-run manually or on a schedule (monthly by default) |
| Cost | 1 to 10 credits per run, depending on history length and how many drivers are analysed. The estimate is shown before you run it |
| Sharing | On Professional and Enterprise, a forecast can be shared through a public link |
API and MCP
- The REST API (Professional and Enterprise) can create, run and read forecasts, their run history and a trust summary.
- The MCP server has forecast tools, so you can ask Claude or Cursor to create and run one.
Forecast a metric from your own data
Forecasting starts on Starter ($99 a month). You can connect your sources and try everything else in the 14-day trial first.