Forecasting — methods and parameters | ATX AnalyticsTable Docs

Forecasting — methods and parametersProjecting a value over time: every forecasting method, seasonality, prediction intervals, scenarios, accuracy measures and every setting. A forecast projects a value over time — sales, volumes, P&L, tickets — the next N days, weeks, months, quarters or years, with shaded prediction intervals. It is available as a dashboard widget, as a chart on a sheet, or as the chart type of any existing chart.

Overview

A forecast projects a value over time — sales, volumes, P&L, tickets — the next N days, weeks, months, quarters or years, with shaded prediction intervals. It is available as a dashboard widget, as a chart on a sheet, or as the chart type of any existing chart.
Add a forecast to a dashboard: Dashboard → Add Widget → Forecast
Add a forecast to a sheet: Reports → Charts → Forecast
Choose the forecasting method: Select the widget → Properties panel → Forecast → Method
1. Open or create a dashboard (Dashboard → Dashboard → New).
2. Dashboard → Add Widget → Forecast places a forecast widget on the canvas.
3. Click the widget to open its properties panel, and bind it to the connection, schema and table that hold the history.
4. In Time series, choose the Date column and the Value column, how a period is combined (Sum by default) and the Period (Day, Week, Month, Quarter or Year). Optionally choose a Forecast each column to draw one forecast per region or product.
5. In Forecast, set Forecast ahead (the horizon) and the Method. Leave Method on Automatic (best backtest) unless you have a reason to force one.
6. In Uncertainty, type up to three prediction-interval levels, e.g. 80, 95.
7. Optionally add scenarios (Upside, Downside) or a management adjustment, and keep Show method and accuracy on to print MAPE, MASE and RMSE under the chart.
On a sheet: select a two-column block "date | amount" before clicking Reports → Charts → Forecast, and both columns are bound at once, with the period guessed from the spacing of the dates.

Available forecasting methods

Automatic (best backtest) — Fits every method the history supports, tests each on the most recent periods it holds back, and keeps the one with the smallest scaled error (MASE).
Naive (last value) — Every future period equals the last complete one. The benchmark any other method has to beat.
Seasonal naive — Every future period equals the same period one season earlier — last March for next March.
Drift — Continues the straight line from the first period to the last.
Moving average — The average of the last few periods, held flat.
Linear trend — A least-squares straight line through the whole history.
Linear trend + seasonality — A straight line plus one seasonal offset per period of the season (each month, each weekday).
Exponential smoothing — A weighted average that leans on recent periods — for a level with no trend.
Holt (trend) — Exponential smoothing of a level and a trend. Damp it to let the trend flatten out.
Holt-Winters (additive) — Level, trend and a seasonal pattern that ADDS the same amount each season.
Holt-Winters (multiplicative) — Level, trend and a seasonal pattern that grows WITH the level (a percentage). Needs positive values.
Holt and the Holt-Winters methods can additionally be damped (Damped trend); the automatic selection tries the damped variants too. A method that needs more history than the data has falls back to the automatic choice, and a "Method changed" chip on the chart says so.

Forecast — projecting a value over time

A forecast answers "where is this heading?": it reads a value over time — sales, volumes, P&L, tickets — and projects it the next N days, weeks, months, quarters or years, with a shaded band for how sure it is. Insert one from Dashboard → Forecast, from Reports → Forecast on a sheet, or pick Forecast as any chart's type; it uses the same sources, title, legend, fonts, number format, PNG, SVG and CSV buttons and PDF reports as every chart, and saves with the project.
It needs three things: the DATE column, the VALUE column and the PERIOD it counts in. Select a block of "date | amount" before clicking the Reports button and both columns are bound at once, with the period guessed from how far apart the dates are. "Combine a period by" adds the rows of a period up (Sum, the default), or takes their average, count, minimum or maximum. The date may be a real date or timestamp, ISO text, a month typed as text such as 2026-03, or European 05.03.2026; the period is worked out in the database on DuckDB, PostgreSQL, SQL Server, Oracle and BigQuery, so only one row per period comes back however large the table is.
"Forecast each" draws one forecast per region or product, the largest five by default with the rest folded into Other, and "Also forecast the total" adds their sum. On a dashboard the slicers, range filters and cross-filters apply BEFORE the periods are counted, so filtering to one region forecasts that region; the chart's own prefilter is added to them. 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.
A period with no rows counts as zero for a sum or a count and is interpolated for an average. The forecast never goes below zero unless the history does. The chart shows the history as a solid line, the forecast dashed in its own shaded zone, and the 80 % and 95 % intervals as two bands; hovering a period lists the actual, the forecast and each band.

Forecast — methods, accuracy, scenarios and parameters

METHOD "Automatic (best backtest)" is the default and the honest one: it fits every method the history can support — naive, seasonal naive, drift, simple exponential smoothing, linear trend, Holt (trend, also damped), linear trend with seasonality and Holt-Winters additive and multiplicative (also damped) — holds back the most recent periods, forecasts them from the rest, and keeps the method with the smallest scaled error (MASE), preferring the simpler one on a tie.
Any method can also be chosen by name; one the history is too short for falls back to the automatic choice and a chip says so. SEASONALITY is detected from the history — two full seasons are needed — or forced on or off; the season length follows the period (12 months, 4 quarters, 52 weeks, 7 days) unless typed. Smoothing parameters are fitted by least squares unless typed. ACCURACY is printed under the chart for every series: the method, MAPE, MASE (below 1 beats the naive benchmark) and RMSE on the held-back periods, the next value, the total of the forecast and its change against the same number of periods before it.
"Draw the test forecast over the history" shows what the method would have said about the periods it was tested on; "Draw the model fit" shows its one-step fit. INTERVALS are prediction intervals at the levels you type (up to three); they widen with the horizon and assume normal, uncorrelated errors, so treat them as a guide rather than a guarantee. SCENARIOS: "Adjust forecast (%)" is a management overlay that moves the forecast and its band, and each scenario adds a line of its own — a flat change and an extra growth per period, such as Upside +10 % or Downside -10 %.
PARAMETERS: the horizon and the adjustment can each be taken from a project parameter, so a parameter read from a cell turns a sheet cell into an input — type 12 in it and the forecast reaches twelve months ahead; a parameter that is not a number leaves the typed value in force and a chip says so. "Learn from" starts the history at a period such as 2024-01 to leave a structural break behind. Changing the method, the horizon, the intervals or a scenario never re-runs the query — only the date, value, period, groups and prefilter do.
CSV exports one row per period and series with the actual, the fit, the forecast, each band, the test forecast and every scenario. Ask Greg can propose a forecast ("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.

Every forecast setting

— Time series —
Date column — Dates, timestamps or month text such as 2026-03. Rows whose date cannot be read are left out.
Value column — The numeric column that is projected.
Combine a period by — How the rows that fall into one period are combined into one value. — Sum; Average; Count of rows; Minimum; Maximum
Period — The calendar period the history is counted in and the forecast is projected in. — Day; Week (ISO); Month; Quarter; Year
Forecast each — Optional category column (region, product…): one forecast per value, in its own colour.
Max series — How many series are drawn; the rest are folded into Other (Sum and Count) or left out (Average, Minimum, Maximum). Default 5.
Also forecast the total — Adds one more series: the sum of all groups, forecast in its own right. Available for Sum and Count only.
— Forecast —
Forecast ahead — Number of periods to project.
Or take it from a parameter — A project parameter — for example one read from a cell — sets the horizon while it holds a number.
Method — The forecasting model. Automatic is the default. — Automatic (best backtest); Naive (last value); Seasonal naive; Drift; Moving average; Linear trend; Linear trend + seasonality; Exponential smoothing; Holt (trend); Holt-Winters (additive); Holt-Winters (multiplicative)
Seasonality — Whether a seasonal pattern is modelled. Detect needs at least two full seasons of history. — Detect; Always; Never
Season length — Periods in one season. Empty = from the period (12 months, 4 quarters, 52 weeks, 7 days).
Damped trend — Shown for trend methods. Lets the trend flatten out over the horizon instead of continuing indefinitely.
Window (periods) — Moving average only: how many of the last periods are averaged.
α (alpha) — Level smoothing, 0–1. Exponential smoothing, Holt and Holt-Winters. Empty = fitted to the history by least squares.
β (beta) — Trend smoothing, 0–1. Holt and Holt-Winters. Empty = fitted.
γ (gamma) — Seasonal smoothing, 0–1. Holt-Winters additive and multiplicative. Empty = fitted.
φ (phi) — Damping factor, 0–1. Shown when Damped trend is on. Empty = fitted.
— Uncertainty —
Prediction intervals (%) — Up to three levels, e.g. 80, 95. Empty = no band. They assume normal, uncorrelated errors.
— History —
Learn from — First period the model learns from, written like 2024-01, 2024-Q1, 2024-W05 or 2024-01-31. Empty = all history.
Leave out the period still in progress — On by default: the period still in progress is drawn hollow and forecast, but the model does not learn from it.
Periods with no rows — How a period with no rows is treated. Automatic counts it as zero for Sum and Count and interpolates it for Average. — Automatic; Count as zero; Interpolate
Never forecast below zero — On by default: the forecast and its band are not drawn below zero unless the history itself goes below zero.
— Scenarios and adjustments —
Adjust forecast (%) — An adjustment moves the base forecast and its band; each scenario adds a line of its own — a flat % change and an extra % of growth per period.
Or from a parameter — Takes the management adjustment from a project parameter, e.g. one read from a sheet cell.
Scenario — One line per scenario; Add scenario creates Upside (+10 %) and then Downside (-10 %) as starting points.
Change % — A flat percentage change applied to the base forecast for this scenario.
Growth %/period — An extra percentage of growth compounded per forecast period for this scenario.
— Accuracy —
Test on the last periods — Periods held back to measure accuracy. Empty = automatic (up to the horizon or one season).
Show method and accuracy — Prints method, MAPE, MASE and RMSE under the chart for every series.
Draw the test forecast over the history — Draws what the method would have forecast for the held-back periods.
Draw the model fit over the history — Draws the model's one-step-ahead fit across the history.
— Prefilter —
Prefilter rows — Keep only rows whose column is one of the values. Added to the dashboard's filters, never instead of them.
— Lines and colours —
Value axis starts at zero — Starts the value axis at zero.
History / Forecast / Band colours — Colours of the actual line, the forecast line and the interval bands.
Max periods fetched — Rows of periods × series the query may return.

Reading the accuracy line

MAPE — Mean absolute percentage error on the held-back periods. Easy to read, but unstable when actual values are close to zero.
MASE — Mean absolute scaled error: the error relative to the naive benchmark. Below 1 means the method beats the benchmark. Automatic selection keeps the method with the smallest MASE, preferring the simpler one on a tie.
RMSE — Root mean squared error, in the units of the value; penalises large misses more than small ones.
next / next N — The forecast for the next period, and the total of the forecast across the horizon, with its change against the same number of periods before it.

Messages on the chart

First {count} rows — The row limit was reached, so the most recent periods may be missing. Raise "Max periods fetched" or choose a longer period.
Parameter not a number — The parameter {names} did not resolve to a number, so the value typed in the settings is used.
"Learn from" not read — "{value}" is not a period of this kind; all history is used.
Top {count} shown — An average, minimum or maximum cannot be added up, so the smaller series are left out rather than folded into Other.
Method changed — {method} needs more history than there is, so the best method that fits was chosen instead.
{period} in progress — The last period is not over yet: it is shown and forecast, but not learned from.
Adjusted {percent} — A management adjustment is applied to the forecast and its band.
{count} rows without a date — These rows have a date the period could not be read from, and are not in the history.

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.