User Guide
Reading the Results page, keeping and reopening runs, changing a forecast.
The Results page
One page, four tabs. It always shows the last run of this browser tab. The top of the page holds:
- the verdict: Ready to use, Usable · n to check or Not usable yet (see Concepts), with one sentence and the run number and duration;
- Change the forecast: back to the wizard, on step 3, with this run's settings and data loaded;
- Download CSV: every series, period, kind (what happened / forecast), value, low, high;
- Download JSON: the results file of the run.
Overview
| Part | Shows |
|---|---|
| Series forecast | how many series have a forecast, out of how many, and how far ahead |
| Typical error | the average error measured on your past, in %; not measured when no test was possible |
| Better than a simple guess | in how many series the method beat repeating the last value |
| Points to check | the checks, red first, then orange, grey, green, each naming its series |
| Forecast chart | one series at a time: what happened, the forecast, the likely range. Series picks it; Show per adds up periods into weeks, months, quarters or years |
Forecast table
The forecast periods with value, low and high, after the most recent history, for the series and grouping chosen above (the two tabs share them). Rows in a tint are forecast. With Show per, each row is a total; (n) is how many original periods it adds up.
Your data
One card per series: history length, data quality, direction over time (no clear trend, a steady trend, a trend that speeds up), seasonal pattern (e.g. repeats every 12 periods), ups and downs (stay about the same size, or grow as the level grows), and the data warnings. Technical details opens the raw profile (mean, standard deviation, stationarity, transform and outlier method chosen, profile reliability).
Explain my data → Ask AI asks PowerAI for a plain-language reading of these patterns.
Quality checks
One card per series:
- method used, its typical error on past data, and x % smaller (or larger) error than the simple guess;
- whether the method gives a low/high range;
- Tests on the errors: leans one way (too low / too high), pattern left unexplained, errors growing with the level. These come with the run's diagnostics; a run reopened without its diagnostics file says so;
- Methods compared: every candidate with its errors (MAPE, RMSE, MAE), the chosen one tagged.
Then, closed by default: Suggestions from the engine (technical) and How the method was chosen (step by step) (the engine's decision log: profiling, transforms, outliers, model choice, with the reason of each).
At the bottom, the AI Forecast Critic reviews the run, can propose a change to the specification (for example another horizon or method), runs the forecast again with it, and compares the result with the previous attempt. You confirm every change, and Restore returns to any earlier attempt.
Keeping and reopening runs
Every forecast you run is kept by PowerForecast: open Forecasts in the sidebar (see the Wizard Guide) and click Results on the one you want. The files saved at the end of a run (and Download JSON / Download CSV) remain a copy you own. To reopen a file: Results → Open downloaded files → choose the results file and the diagnostics file together (the profiling and specification files are accepted too). Pasting the JSON works as well.
Upload in the sidebar loads a forecast specification (settings + data) into the wizard, to run it again or change it. A results file is not a specification: reopen it in Results.
Changing a forecast
- Change the forecast on the Results page reopens the wizard on step 3 with the run's settings and data.
- To forecast with new actuals: replace the file on step 2 (or load an updated workbook on step 1) and run again.
- To compare two runs: download the CSV of each and compare them in your spreadsheet. The demo US retail sales does this with a budget forecast and its re-forecast.
Tips
- More history, better seasons. For monthly data, three years or more lets the engine learn the yearly pattern and test it.
- One column per comparable thing. Very different products in one series blur each other's patterns.
- Forecast what you plan. A 12-month budget needs 12 months, not 36.
- Re-forecast when actuals arrive. A forecast is sharpest at its start; refreshing it monthly or quarterly keeps it close to reality.
- Read orange points before using the numbers, especially tends to be too low / too high: they tell you which way to adjust.