Predictive forecasting (premium forecasting)

What it does

Premium forecasting adds a machine-learned Prediction column to the forecast grid: a projected revenue figure for each row, derived from closed-deal history and the current open pipeline. An optional prediction factors view shows the top influences behind each row's number.

Key facts

Requires a Dynamics 365 Sales licence that includes premium forecasting — this is a Sales Premium capability, not part of Sales Enterprise. The Prediction column type only appears in the Layout step when the tenant is licensed for it.
Available only where the forecast's rollup entity is Opportunity. It works with any hierarchy on that basis.
Not available in Government Community Cloud, France or India.
Minimum viable training data is more than ten closed opportunities carrying all four of Actual Value, Actual Close Date, Est. Revenue and Est. Close Date. Ten is the floor, not a target — accuracy climbs with volume.
Open opportunities need Est. Revenue and Est. Close Date populated, since those are what the model projects forward.
Every underlying opportunity must be owned by a system user. Team-owned or otherwise non-user-owned opportunities produce no prediction.
The model is selected automatically from the available data. There is no algorithm choice, no training schedule to configure and no model management UI.
After activating a forecast with a Prediction column, values appear in roughly two hours.
The Prediction column is excluded from normal forecast recalculation. It refreshes on its own cycle, every seven days.
Enable prediction factors is a toggle on the Advanced step of forecast configuration. It surfaces the top contributing factors per row.
Filters applied to the forecast do not degrade the model — the prediction is built from the data regardless of what the grid is showing.
Turning on predictive opportunity scoring improves the quality of the underlying model, and is worth doing alongside.

When to use / skip

Use it where the client already has two or three years of clean closed-won and closed-lost history in Dynamics, sells on a repeatable motion, and has a leadership team that will actually look at the number. In that setting it is a genuinely useful sanity check against what the sales hierarchy is committing.

Skip it on a fresh implementation. A client who migrated opportunities last month has no history, and the prediction will be confidently wrong in a way that destroys trust in the whole forecasting build. Get twelve months of real usage first, then revisit.

Skip it also where deal volume is low and value is lumpy — enterprise sales with twenty deals a year. The model needs volume, and the client's own judgement will beat it.

And be straight about the licensing. Premium forecasting is the kind of feature that gets demoed on Enterprise licences and then discovered at deployment. If the client is not on Sales Premium, the Prediction column is not there and no amount of configuration will produce it.

Configuration decisions

Whether the client is on, or willing to move to, Sales Premium — before this appears in any design document.
Whether prediction factors are exposed to sellers or only to managers. Factors invite challenge, which is healthy but noisy.
How the predicted number sits next to Committed in the grid, and which one leadership is told to read.
Whether to clean historical opportunity data (actual values and close dates on old closed records) as a precursor, and who pays for that.
Whether predictive opportunity scoring goes in at the same time, given it improves the model.
Which forecast configurations carry a Prediction column — it is per-forecast, and adding it everywhere is rarely useful.

Gotchas

Migrated or test data poisons the model. Opportunities with a created date after their close date, or bulk-loaded records with default values, are exactly the pattern the docs warn about, and migrations produce both.
Team-owned opportunities silently produce no prediction. At clients who own deals by team for access reasons, this can wipe out most of the grid with no error message.
The seven-day refresh cycle means the prediction will not move when the pipeline does. Managers assume it is broken. It is not.
Prediction columns cannot be used inside an adjustable calculated column, so you cannot build "predicted total" as an editable field.
France and India tenants get no premium forecasting. For a multinational client this can mean the feature works for some regions and not others.
The two-hour wait after activation means the first thing the client sees is an empty column. Set that expectation before the demo.

Consultant notes

Never demo predictive forecasting on sample data. It will look plausible in the demo and nothing like that in production, and you will own the gap.
Run a data readiness check first: count closed opportunities with all four required fields populated, and check owner types. Ten minutes of FetchXML saves a fortnight of argument.
Position it as a second opinion on the committed number, not as a replacement for the sales team's judgement. Clients who expect it to replace the forecast call are set up to be disappointed.
If the client wants it and the data is not there, phase it — build the forecast now, add the Prediction column at the next release once real history exists.
Check the licence entitlement in the admin centre yourself rather than taking the client's word for what they bought.

Worth another look if the seven-day prediction refresh shortens, or if the regional exclusions change.