This FAQ addresses common questions and clarifications that arise while working with Forecasting. It covers rolling forecast concepts, statistical forecasting configuration, closing and extending the forecast horizon, reforecasting open periods, and writing the finalized forecast to the Fabric SQL database.
A closed period reads from an existing measure — typically actuals — so the forecast reflects what actually happened. An open period accepts data input, either manual entry or values copied from another measure during period setup. As time moves forward, you progressively close periods to replace forecast values with actuals, keeping the full-year view accurate.
Yes — you can create multiple forecast measures and display them as separate columns on the same sheet. Each forecast measure has its own open and closed period configuration.
Yes — go to Planning > Manage Measures to delete the forecast measure. This removes the measure and all associated open period configurations from the planning sheet.
The algorithm needs sufficient data to reliably detect seasonal patterns. With less than two years of data, the model cannot distinguish a true seasonal pattern — for example, a Q4 uplift every year — from a one-time event. Two years give the model one full seasonal cycle to learn from and one to validate against.
The confidence interval defines the range within which the algorithm expects the true value to fall, given the variability in historical data. A 90% confidence interval means the algorithm is 90% confident that the actual value will fall within the shaded range shown in the forecast preview. A higher confidence produces a wider range; a lower one produces a narrower range. The point estimate — the central line — does not change; only the uncertainty band changes.
Yes — the Growth Factor applies an additional uplift on top of the model's statistically derived values. Setting it to 4% instructs Predict to grow its output by 4% before saving, reflecting a business assumption about market growth that the historical data alone wouldn't capture.
Bottom-Up calculates the forecast at the most granular level first and aggregates up to the highest dimension category and total. Top-Down calculates at the top level and distributes down. Bottom-Up is generally more accurate for granular sales data because each child category's seasonality pattern is captured independently. Top-Down is faster but smooths over brand-level differences.
Set Seasonality tells the model which cycle lengths to detect and account for when generating the forecast. Select Year to instruct the model to look for patterns that repeat annually — for example, a consistent Q4 uplift each year. Select Quarter to add a second layer, capturing patterns that repeat within the year at the quarterly level.
Historical data is shown in grey, while predicted values are shown in green, with the confidence range displayed as green shading in the background. This allows you to visually assess whether the model's output looks reasonable before committing the values to the forecast measure.
Yes — you can re-run Predict on the same measure with different settings. The new values will overwrite the previously saved forecast values for the selected date range.
Yes — you can filter the rows before running Predict, or configure the row selection in the Predict dialog to target a specific subset. The Bottom-Up evaluation will then run only for the rows in scope.
The closed month is locked — it populates automatically with the linked measure and turns grey to indicate it cannot be edited. Any forecast assumption previously entered for that month is replaced by the actual value.
No — closed period values cannot be edited. They will be greyed out, indicating that those values are non-editable.
Overwrite forecasts replace the closed period forecast values with actual performance data — this is the default and the most common choice for a rolling forecast, since you want actuals to be the single source of truth for closed periods. Retain forecasts keeps the original forecast values and only fills in actuals where forecast values are blank. Use Retain when you want to preserve the forecast for variance analysis alongside actuals.
Yes — closing a period replaces the forecast value with the actual value, so the full-year total will reflect the actual figure. This is expected and correct behavior. The variance between forecast and actuals is visible when you show both measures side by side.
You can close a custom range — not just the previous month. In the Close Period dialog, select Custom under Close Period Till and specify the end date of the range you want to close. This is useful if you are catching up after a gap or closing an entire quarter at once.
Extend Forecast Range automatically adds one period to the end of the forecast horizon when you close a period. This is what makes it a true rolling forecast — for example, closing January 2026 and extending by one month adds January 2027, so the forecast always covers the same forward-looking window.
The period closes, and actuals populate correctly, but the forecast horizon shrinks by one month. You would need to manually extend it later using Reforecast.
Reopening a closed period is not a standard operation — it would require adjusting the close period configuration. In practice, actuals corrections are typically handled in the source system, and the close period is re-run.
Reforecast opens a new period setup dialog that lets you reconfigure the open period behavior — for example, changing the copy source or growth assumptions — and then redistributes values across the remaining open periods accordingly. Editing cells directly is faster for a single adjustment, but does not recalculate the full distribution. Use Reforecast when assumptions change broadly; use direct editing for targeted corrections.
It maps each target month to its corresponding source month within the defined range. It's a positional copy — the first month of the target range gets the value from the first month of the source range, and so on.
Proportionally — each child category within a parent category receives the growth distributed according to its share of the category total, so the existing distribution is preserved. If it needs to be distributed based on any other measure or equally, this can be configured after entering values in the cells.
