Outcome record
Monthly demand forecast, timed on a held-out quarter
In May 2026, nine learners on the Spreadsheet to Forecast evening cohort produced a monthly pack twice on the same anonymised series: once with their current spreadsheet habit, once with the class cleaning checklist, basic model, and sanity checks. Mean absolute percentage error on the held-out quarter moved from 18.4% to 11.2%.
Course details All outcome records
What the sample looked like
The file was a 36-month series for twelve product lines, with the last three months held out. Units were mixed on purpose: two lines in cases, the rest in pieces. Two lines had missing months. One line had a duplicate SKU that inflated history if a learner summed without checking. Calendar Chinese New Year effects sat in the commentary notes. They were absent from any helper column, so they had to be found by reading.
The series is fictional and anonymised. It was built from patterns we see in scoping calls with small wholesale and clinic-style operations in Singapore. It is still a classroom file. A live hierarchy with a mid-year rebrand will behave worse than this pack.
Protocol on the day
Each learner produced a pack: cleaned table, forecast for the held-out quarter, a short narrative, and a chart. MAPE was computed by the second marker from the learner’s submitted quarter against the hidden actuals. Learners did not see the actuals until after both passes. Time to produce the pack was recorded as a secondary measure; the published headline is error, because that is what a manager can audit.
The rebuild used a transparent method. Learners chose, against the sample, between a moving average and a simple seasonal naïve, and had to write the choice in the narrative. The assistant was allowed to propose a method and was not allowed to hide it. Sanity checks flagged any line more than 30% away from the same quarter last year, forcing a comment.
The numbers
| Measure | Baseline | After |
|---|---|---|
| MAPE on held-out quarter | 18.4% | 11.2% |
| Median time to complete the pack | 84 min | 51 min |
| Learners kept | 9 | |
| Date | May 2026, hybrid evening cohort | |
MAPE is an unweighted mean across the twelve lines. A volume-weighted error would have looked kinder, because the two worst lines were small. We kept the unweighted figure because a manager still has to explain every line. One learner’s after MAPE was slightly worse than baseline after they over-corrected a seasonal factor; that row stays in the nine.
What learners built
The cleaning checklist was one page: missing months, unit mixes, duplicate SKUs, and a forced plot of each line before modelling. The forecast sheet exposed the method in a labelled row, so a colleague could reproduce the number without opening the assistant. The narrative template had three headings: what changed, what was assumed, what a reader should not conclude.
The assistant drafted the narrative after the numbers were locked. When a learner let the assistant propose numbers first, the second marker could see it: the sanity-check comments were empty. That pattern showed up in two baseline packs and in none of the after packs.
Limits that belong next to 11.2%
Nine people is a small evening class. Hybrid delivery means two of the nine were remote; their files were scored the same way, and their MAPE sat inside the same spread. The held-out quarter was three months, which is short. A live business with promotions, stock-outs, and a sales team that pads orders will need a different error conversation than MAPE on a tidy demo.
We did not measure whether the narrative was believed by a real manager. We measured whether the numbers could be audited and whether MAPE moved. Persistence after May 2026 is unknown. If you need a model that writes to a warehouse, this programme is the wrong object; the course stays in the spreadsheet you already use.
Measure this on your own series
Bring a year of monthly actuals with names stripped. Agree the error formula before the first evening. If the series cannot leave the building, run the cohort on-site. The public record remains this May 2026 classroom file. Programme page: Spreadsheet to Forecast.