Forecasting — Methods, Scenarios, Accuracy

Forecasting — where the numbers are headingA value over time, projected the next N periods with honest bands — fitted live from the database, the method chosen by backtest, scenarios drawn beside it. A forecast answers the question every monthly pack eventually asks: where is this heading? It reads a value over time — revenue, volumes, tickets, P&L — and projects it the next days, weeks, months, quarters or years, with a shaded band for how sure it is. It is not a separate tool: it is a chart type, inserted from Dashboard → Add Widget → Forecast, from Reports → Charts → Forecast on a sheet, or picked as the type of any existing chart, and it keeps the same sources, number formats, PNG, SVG and CSV buttons and PDF reports as every chart. The database is asked one question — the value per period, and per group when there are groups — so ten thousand orders or a hundred million cost the same; everything after that, fitting, backtesting, intervals and scenarios, happens in the app, which is why changing the method or the horizon is instant. Every screen in this chapter comes from the Forecast dashboard of the AnalyticsTable-Showcase project that ships with the download, so you can open the same board and change the same settings.

What it does

Three inputs and a forecast: the DATE column, the VALUE column and the PERIOD — day, ISO week, month, quarter or year
Rows of one period combined by Sum (the default), Average, Count of rows, Minimum or Maximum — counted in the database on DuckDB, PostgreSQL, SQL Server, Oracle and BigQuery
Eleven methods: naive, seasonal naive, drift, moving average, linear trend, linear trend + seasonality, exponential smoothing, Holt, Holt-Winters additive and multiplicative — Holt and Holt-Winters also damped
Automatic (best backtest) is the default: every method the history supports is fitted, tested on the most recent periods it holds back, and the one with the smallest scaled error (MASE) is kept
Seasonality detected from the history (two full seasons needed) or forced on or off; the season length follows the period — 12 months, 4 quarters, 52 weeks, 7 days — unless typed
Up to three prediction-interval levels, drawn as nested bands that widen with the horizon
Forecast each: one forecast per region, product or channel, the largest five by default with the rest folded into Other, plus the total forecast in its own right
Scenarios — a flat % change and an extra % of growth per period each — and a management overlay that moves the base forecast and its band
The horizon and the overlay can come from a project parameter, so a cell or Settings → Parameters drives the forecast
Method, MAPE, MASE, RMSE, the next value, the total of the horizon and its change against the same span before, printed under the chart for every series
Dashboard slicers, range filters and cross-filters apply before the periods are counted; the chart's own prefilter is added to them, never instead of them
CSV export of every period and series — actual, fit, forecast, each band, the test forecast and every scenario

How it works

Add the widget (Dashboard → Add Widget → Forecast) and its properties open. Pick the table, then the Date column and the Value column — the widget shows what it still needs until both are set. On a sheet, select a two-column block "date | amount" before clicking Reports → Charts → Forecast and both are bound at once, with the period guessed from how far apart the dates are.
The date may be a real date or timestamp, ISO text, a month typed as 2026-03 or a European 05.03.2026. Rows whose date cannot be read are left out and counted in a chip, never silently turned into a period.
The database returns one row per period (and per group). A period with no rows counts as zero for a sum or a count and is interpolated for an average; Periods with no rows overrides that.
The last period, when it is still in progress — this month is not over yet — is drawn as a hollow point and forecast, but the model does not learn from a half month. Leave out the period still in progress turns that off.
On Automatic, the app holds back the most recent periods (up to the horizon or one season — Test on the last periods overrides it), forecasts them from the rest with every eligible method, and keeps the one whose error, scaled against the naive benchmark, is smallest. A tie goes to the simpler method.
A method named by hand that needs more history than there is — Holt-Winters on 21 months, for instance — falls back to the automatic choice, and a "Method changed" chip says so. The chart never pretends to have used a model it could not fit.
Smoothing parameters α, β, γ and φ are fitted by least squares unless typed. Intervals assume normal, uncorrelated errors and widen with the horizon — a guide to the spread, not a guarantee.
Changing the method, the horizon, the intervals or a scenario never re-runs the query; only the date, value, period, groups and filters do. That is what lets you click through methods on a large table without waiting.
Ask Greg can propose one — "forecast revenue by month for the next 6 months, by region" — and certifies it by the history it reads, the total and the number of periods, never by the future, which no check can prove.

Practices that keep it reliable

Leave Method on Automatic unless you have a reason; when you do force one, keep Show method and accuracy on and compare its MASE with the automatic choice
Read MASE before MAPE: below 1 beats the naive benchmark, and MAPE is unstable when values come close to zero
Turn on Draw the test forecast over the history when presenting a forecast — it shows the audience what the method would have said about months they already know
Use Learn from to leave a structural break behind — a reorganisation, a new product line, a pandemic year — rather than letting it bend the trend
Keep scenarios and the management overlay visibly labelled; an overlay moves the band too, and the chip on the chart is the reader's only cue
Forecast the total in its own right (Also forecast the total) rather than adding up group forecasts — the sum of five forecasts is not the forecast of the sum

In a regulated environment

A forecast is a saved chart definition in the project file — method, horizon, intervals, scenarios and filters are a reviewable text diff in Git, not a spreadsheet someone re-ran by hand.
Every figure is refitted from the source on refresh, so the board pack's outlook cannot drift from the numbers it was built on.
The accuracy line documents the model choice on the chart itself — which method, and how it scored on held-back periods — the evidence a model-risk or finance reviewer asks for first.
Ask Greg certifies a proposed forecast by the history it reads, never by its projection; an AI-built chart therefore carries the same checkable provenance as one built by hand.

Every option, on a real screen

Automatic method, with the model fit and the test forecast drawn (Properties → Forecast → Method). The default and the honest choice. Every method the history supports is fitted and tested on the most recent periods it holds back; the one with the smallest scaled error is kept — here damped Holt. The dotted line is the model's one-step fit across the history, the violet one what it would have forecast for the held-back months. If those two lines miss the actuals badly, so will the forecast.
Naive — the benchmark every method has to beat (Properties → Forecast → Method). Every future period equals the last complete one. It is rarely the forecast you publish, but it is the yardstick: MASE divides a method's error by the naive error, so a MASE below 1 means "better than doing nothing clever". Notice the band — with no trend to lean on, it fans out fastest.
Daily data with a weekday pattern (Properties → Time series → Period). Switch the period to Day and the season length becomes 7: a linear trend plus one offset per weekday reproduces the weekly rhythm of orders. Learn from starts the history at 1 June 2026 so four months of days are fitted, and 28 periods ahead is four weeks.
Quarters and a moving average (Properties → Forecast → Method). Quarterly periods smooth the noise of single months. The moving average holds the mean of the last few periods flat — Window sets how many — which suits a level without a trend, and makes a sober baseline for a planning discussion.
When a method needs more history than there is (Properties → Forecast → Method). Holt-Winters models a seasonal pattern and needs two full seasons — 24 months — to fit one. The Showcase holds 21. Rather than fit nonsense, the chart falls back to the best method that the history supports and says so with a "Method changed" chip; the accuracy line names the method actually used.
One forecast per product, plus the total (Properties → Time series → Forecast each). Forecast each draws one series per value of a category column, each with its own automatically chosen method. Max series keeps the largest few and folds the rest into Other; Also forecast the total adds the sum, forecast in its own right rather than by adding up the parts.
A management overlay and growth scenarios (Properties → Scenarios and adjustments). Adjust forecast moves the base forecast and its band by a flat percentage — the overlay finance applies after the review meeting — and the chart wears an "Adjusted" chip so nobody mistakes it for the model. Each scenario adds a line of its own: a flat change, an extra growth per period, or both.
Three interval levels and a shorter history (Properties → Uncertainty → Prediction intervals (%)). Type up to three levels and they nest — 50 % inside 80 % inside 95 %. Learn from 2025-07 drops the first half of 2025 from the fit; with less history and nine periods ahead, the bands are visibly wider. That is the point of showing them.
A prefilter fixed on the chart (Properties → Prefilter → Add filter). The chart's own prefilter keeps only rows whose column holds one of the listed values — here Europe and the Aurora Laptop — whatever the dashboard's slicers say. It is added to the slicers, never instead of them, so choosing a channel above narrows this chart further.
Volumes with a damped trend (Properties → Forecast → Damped trend). Any numeric column can be forecast — units, tickets, headcount. Damped trend lets Holt's trend flatten over the horizon instead of continuing for ever, the safer assumption for anything that cannot grow without limit.

Screens

The Showcase's Forecast dashboard — slicers on top filter the history every forecast learns from; the revenue outlook below carries 80 % and 95 % bands, a plan and a downside.
Monthly revenue, six months ahead (the horizon is the forecastMonths parameter). Automatic chose damped Holt: MASE 0.41, so it beats the naive benchmark by more than half.
Forecast each = region: four forecasts, each with its own automatically chosen method and accuracy line.
Units by channel — the value does not have to be money. Damped Holt lets each trend flatten over the horizon.
Average margin per order — Combine a period by Average, and the value axis free of zero so a narrow range stays readable.
Properties, top half — Time series (date, value, combine, period, forecast each), Forecast (horizon or parameter, method, seasonality, damping) and Uncertainty.
Properties, bottom half — Scenarios and adjustments, Accuracy (holdout, accuracy line, test forecast, model fit), Prefilter, and Lines and colours.

Glossary terms on this page

KPI dashboard: definition

ATX Institute on LinkedIn

This site needs JavaScript for the full experience. Every page above is reachable without it.