PW Intelligence · Forecast report

The admissions forecast — accuracy, and a correction

Two things in one page: how accurate the engine has been (measured, not claimed), and a fix we just shipped so the end-of-cycle number stops over-counting a tail that last year barely existed.

As of 13 Aug 202616 weekly replays · Apr 20 → Aug 3 SOT · School of Technology
1.0%
error on today's admissions
end-of-day · n=16
6.0%
error on the month's total
end-of-month · n=15
605
projected admissions, cycle end
was 662 · range 543–697
26→1
October fix (last year: 1)
tail over-count removed

In one line. Near-term the forecast is tight and trustworthy — it calls the daily admissions count to within ~2 students and the month to within 6%. But its end-of-cycle number was inflated by a tail bug: it projected 26 admissions for October when last year all of October produced one. Fixing it drops the cycle projection from 662 to 605 and makes the checkpoint ladder consistent. Details below.

What changed

The tail was counting students who never enrol

The engine's pipeline method spreads students still waiting in the funnel across future months using last year's conversion-lag curve. Late in the cycle that inventory is mostly dead — shortlisted students who were never going to pay — but the method kept projecting them forward, dumping 64 phantom admissions into a September–October window where last year saw 45. We grounded the tail to last year's real monthly shape. Near-term projections (where the pipeline is genuinely predictive) are unchanged.

Monthly admissions — what the engine projected before vs after the fix, against last year's actuals
MonthBeforeAfterLast year (actual)Reads as
August13213289unchanged
September865444realistic
October2611fixed
End of cycle662605635−57
When does the cycle actually end? Last year the last SOT admission landed on 11 October, but the cycle is effectively over by end of September — the monthly count fell 202 → 197 → 89 → 44 → 1. So a realistic forecast must let the tail fall to near-zero, which is exactly what the fix restores.

Where it stands today

471 booked → 605 by cycle close, on a consistent ladder

The checkpoint ladder now ties to the same month-by-month engine that produces the cycle number — so the steps add up. Previously end-of-month said +35 while end-of-cycle said +197; that gap was an artefact of two different methods, now reconciled.

CheckpointAdmissions+ from nowRegistrations+ from now
Now (13 Aug)471—7,704—
End of day473+27,724+20
End of week (16 Aug)479+87,807+103
End of month (31 Aug)550+798,401+697
End of cycle605+1348,779+1,075

End-of-month → end-of-cycle is now just +55 (September 54 + October 1) — matching last year's Sep–Oct tail of 45. No more phantom jump.

Interactive · hover the line

The end-of-cycle call, week by week

Each point is what the engine expected the full-cycle total to be, using only data knowable that day. The corrected line (teal) settles near 605; the old method (dashed) kept climbing past 660 as it over-counted the tail. Hover any week for the exact numbers.

Corrected end-of-cycle call Old (over-counted tail) Engine low–high band Booked to date

Interactive · hover a month

The season is a wave, not a plateau

Admissions build to a June–July peak and taper fast. The corrected forecast (teal) sits just above last year (grey) through the peak — the cycle is running hot — then converges to last year's near-zero tail. This shape is why a large September/October was never realistic.

Last year actual This year (actual + forecast) Apr–Jul = booked · Aug–Oct = projected

Track record

Accuracy by horizon

WAPE = weighted absolute percent error, scored against what actually happened. Lower is better. The verdict says how far to trust each horizon.

Admissions — the number leadership tracks

HorizonScoredAvg missWAPEVerdict
End of day161.71.0%near-exact
End of week1613.06.8%reliable
End of month1512.86.0%reliable
1 month ahead1137.525.0%directional
2 months ahead650.528.9%directional
3 months ahead261.033.0%rough
Why is the month-ahead error high — and will the fix lower it? Honestly, no — because it's a different problem. The month-ahead misses were under-shoots from early in the cycle (April–June predictions came in low because the season ran hotter than last year). This fix corrects late-cycle over-shooting of the tail. They're opposite. The month-ahead error shrinks on its own as the cycle matures and less runway remains for surprises.

What to do with it

Actionable read

1
Plan against the near term as hard numbers. Today / this week / this month admissions are inside ~6% — safe for staffing, seat-blocking, and weekly targets. Use 550 by 31 Aug.
2
Use 605 as the cycle number, but watch the top of the band. The season has out-run last year all year; the realistic range is 543–697, and the upper half is live if August keeps its pace.
3
The gap to 1,000 is structural, not a pacing miss. At 605 projected, closing to target needs volume — more registrations now — not a conversion tweak. September is the last month that can move the number.
4
Don't quote a month-ahead registration figure. That's the one output the backtest says to withhold (73%+ error); lead with the admissions projection instead.

Spot-check

The analyst's answers hold up

Two recent chatbot responses were checked against the raw data:

Bottom line

Trust it up close; the tail is now honest