clariBI Forecasting: backtested projections, prediction ranges and drivers

Forecasting in clariBI projects a metric forward, such as weekly revenue, daily sign-ups or monthly MRR, and tells you how sure it is. It also shows which of your other metrics tend to move with it, flags unusual points in the history and dates the places where the trend changed.

The forecast itself is computed with statistics libraries (numpy, pandas and scikit-learn). No language model produces the numbers, so the same data and settings give the same result. A language model writes only the short summary that describes the result. Forecasting is included on the Starter, Professional and Enterprise plans. It is not part of the trial.

Why not one method for everything

Many tools apply a single method, often exponential smoothing, to every series. That works for series that suit the method and quietly fails on the rest. A daily sign-up count with a weekly rhythm, a lumpy B2B revenue line and a steadily growing MRR series need different treatment.

clariBI tests several methods on each series and keeps the one that would have predicted your own history best.

How a forecast is made

1. Test on the past. clariBI hides the most recent part of the history, forecasts it from the rest and compares the result with what actually happened. It repeats this several times, moving forward each time. This is a walk-forward backtest, and the error is measured as sMAPE (symmetric mean absolute percentage error).

2. Keep the best method. The candidates range from simple to more involved:

• naive ("same as last period") and seasonal naive ("same as the same day last week")
• moving average, drift and linear trend
• autoregressive models
• Holt-Winters (additive, multiplicative and damped)
• Theta
• gradient boosting
• models that use your other metrics as inputs, and a multi-series model
• an average of the best three

The method with the lowest backtest error on your data is used, and the result says which one won. Simple methods win more often than people expect. If "same as last week" beats everything else, that is worth knowing too.

3. Show the uncertainty. The errors from those backtests set the ranges around the projection: the bands that held 50, 80 and 95% of past outcomes. A tight band means the metric has been predictable. A wide band means it has not, and planning should allow for that.

Drivers: which other metrics move with it

clariBI compares the forecast metric with the other metrics in your account using Pearson and Spearman correlation, including at lags of up to 14 days. A metric that tends to move a week before your target shows up as a leading indicator even if its same-day correlation is weak.

Testing many pairs produces some matches by chance. clariBI applies a Benjamini-Hochberg correction that holds the false discovery rate at 5%, so fewer coincidences reach the list. Each remaining driver is then re-tested on resampled data (a bootstrap); it is marked stable only if the relationship holds up, and fragile otherwise.

A driver is a place to look, not a cause. If trial sign-ups lead paid conversions by 10 days, that is useful for planning. It does not prove that more sign-ups will produce more revenue.

Anomalies and changepoints

• Anomalies are points far outside the normal range for the series, found with z-scores and an interquartile range fence, and labelled by severity. They are marked on the date they happened. Anomaly detection for business metrics explains both methods.
• Changepoints are dates where the trend changed slope, found with segmented regression and an F-test. The result shows the rate before and after, so when a price change or a new channel altered the trend, you can put a date on it.

What you can forecast

A forecast is attached to one metric, which can come from four places:

• a connected source's time series, such as Stripe revenue, Google Ads spend or GA4 sessions
• a dashboard widget, or every forecastable widget at once with Forecast Dashboard
• a metric in a report, where each regeneration adds a data point
• a goal's progress, using Forecast value or Forecast progress % on the goal

Granularity and horizon

Series at 5, 15, 30 and 45-minute intervals are also supported. The longest horizon is not the recommended one. Forecasts get less certain the further out they go, and the bands widen accordingly. For a volatile series, keep the horizon short.

Forecasts can be re-run on demand or on a schedule (monthly by default). clariBI checks for due runs every 30 minutes.

What it costs

Each run costs 1 to 10 AI credits, by a fixed formula:

• 1 credit for the run
• plus 2 if the method for this series has to be backtested again (a repeat run of a series clariBI has already characterized skips this)
• plus 1 for every 4 drivers found, rounded up
• plus 1 or 2 for a long history, with thresholds that depend on the granularity

The estimate is shown before you run it. If a run fails, the reserved credits are refunded. A scheduled forecast that fails three times in a row, for example because the organization is out of credits, pauses itself. More on credits in how AI credits work.

Judging a forecast

Every forecast has a Trust view with the backtest error, the method used and, as real values arrive, the error the forecast actually achieved over time. Read it before you put a projection in a board deck. The help article on how to judge an answer covers the Trust view alongside the other checks.

From code and from AI assistants

• The REST API (Professional and Enterprise) can create, run and read forecasts, their run history and the trust summary, and estimate a run's cost. See the REST API reference.
• The clariBI MCP server has forecast tools, so you can ask Claude or Cursor to create a forecast, run it and read the result.
• On Professional and Enterprise, a forecast can be shared through a public link.

What it is not

Forecasting projects time series and surfaces statistical relationships. It is not a causal model, it cannot fill gaps in data you do not have, and it does not know about the price rise you are planning next month. Use it to set realistic targets and to see early when a metric drifts off course. Revenue forecasting for subscription businesses shows how to combine a statistical forecast with the plans it cannot see.

To try it, connect Stripe or your sales data on a Starter plan, open Forecasting, pick weekly revenue with a 12-week horizon, and look at which method won and how wide the 80% band is. The forecasting feature page has the full reference.