Concepts
The ideas behind PowerForecast, from the data you bring to the verdict you read.
Series and periods
A series is one line of history measured at regular periods: monthly sales of one department, weekly claims of one branch, daily calls of one queue. A forecast can hold many series. Each one is studied and forecast on its own, with the method that suits it.
The time unit (hourly, daily, weekly, monthly, quarterly, yearly) is read from the gaps between your dates. Every series of one forecast shares it. The data step warns when:
- the gaps vary (missing or duplicated periods),
- series have different lengths,
- a date appears twice in a series (keep one row per date),
- a series is short for its time unit (about 48 hourly, 30 daily, 12 weekly, 24 monthly, 8 quarterly or 5 yearly values make a reliable minimum).
Horizon
How far ahead: the number of periods to forecast. Uncertainty grows with distance. The wizard warns when the horizon is more than a third of the shortest history: the last periods will be less reliable.
What the engine does with each series
- Profiling. Several statistical tests look for a trend (steady, or speeding up), a seasonal pattern (and its length, e.g. 12 months), whether the ups and downs grow with the level, and the overall data quality. A pattern counts only when several methods agree; when the history is long enough, the pattern is also confirmed by checking that taking it into account really improves a forecast of held-out periods.
- Preparation. Depending on the profile, the series may be transformed (e.g. a logarithm when swings grow with the level) and outliers handled (detected in what remains after removing trend and seasons, so a real Christmas peak is not treated as an error).
- Choice of method. A shortlist of methods fitting the profile is tested on your past by cross-validation: the method forecasts periods it has not seen, several times, and its errors are measured. The method with the smallest error becomes the champion. The simplest method, Naive (repeat the last value), is measured too, as the baseline.
- Forecast. The champion forecasts the horizon, with a likely range when the method provides one.
- Diagnostics. The champion's errors on the past are tested: do they lean one way (bias)? is there pattern left in them? do they grow with the level?
Methods in the engine: AutoARIMA, ETS (exponential smoothing), Theta, MSTL (several seasons), SeasonalNaive and Naive. With Automatic (the default), the engine chooses per series. With Manual, you pick the candidates among AutoARIMA, ETS, SeasonalNaive, Theta and MSTL. Naive stays the baseline every method is measured against.
The numbers on the Results page
| Number | Meaning | Basis |
|---|---|---|
| Typical error (MAPE) | average gap between forecast and reality, in % of reality | cross-validation on your past; not measured when the history was too short |
| Better than a simple guess | the champion's error is smaller than Naive's | same cross-validation |
| Smaller / larger error than a simple guess | (Naive error − champion error) / Naive error | RMSE of both in cross-validation |
| Likely range | low – high band around each forecast value | from the method; its width follows the likely range setting (default 95 %) |
The page shows a number only when the engine measured it. A series that could not be tested says so. Totals per week, month, quarter or year (Show per) add up the forecast values. They show no range, because adding the bounds of each period does not give a valid range for the total.
The verdict and the checks
Each check is computed from the run and shown in plain words, with the series it concerns:
| Level | Check | Fires when |
|---|---|---|
| red | No forecast | a series got no forecast value |
| orange | Not better than a simple guess | the champion's error is larger than Naive's |
| orange | Large typical error | typical error above 20 % |
| orange | Data quality is fair or poor | the profile rates the data fair or poor |
| orange | Tends to be too low / too high | the past errors lean one way (bias test) |
| grey | Not tested on past data | no cross-validation possible (history too short) |
| grey | No low/high range | the method gives no interval |
| grey | Range less reliable | pattern left in the errors, or errors growing with the level |
| grey | Notes about your data | warnings from profiling |
| green | Typically off by x % · Better than a simple guess | the reassuring facts |
Verdict: red if any red check, otherwise orange if any orange check, otherwise green.
Settings
| Setting | Default | Effect |
|---|---|---|
| What to forecast | — | a label for the run (demand, workload, revenue, claims, transactions, or your own word) |
| Split by | — | a label (per product, region, segment, branch, category, or over time); the series themselves come from your data |
| Time unit | from the data | change only if the detection was wrong |
| Width of the likely range | 95 % | 80 %, 90 %, 95 %, 99 % |
| Scenario | baseline | a label carried with the run (optimistic, pessimistic, stress) |
| Method | Automatic | or a list of methods to try |
Next: the Wizard Guide.