PW Intelligence · Forecast validation

How accurate has the admissions forecast been?

The forecasting engine, replayed as-of every week of the cycle and scored against what actually happened since. This is the honest track record — not a claim, a measurement.

Through 12 Aug 2026 16 weekly as-of points · Apr 20 → Aug 3 132 scored predictions

1.0%

error on today's admissions count

end-of-day WAPE · n=16

6.4%

error on the month's admissions

end-of-month WAPE · n=15

±2

avg. miss on daily admissions

mean abs. error · admissions

25%

error a full month ahead

month+1 WAPE · admissions

The short version. For the horizons people actually plan against — today, this week, this month — the admissions forecast is tight and reliable: it has called the running admissions count to within ~2 students on any given day, and the month's total to within ~6%. It is deliberately more cautious the further out it reaches: a full month ahead it is directional (±25%), and it consistently under-shot the far horizon because this cycle ran hotter than last year's seasonal shape. Registration volume a month out is its known weak spot. Read the near term as a number; read the far term as a direction.

Where it stands today

466 admissions booked → 662 projected by cycle close

Admissions

466 → 662

end-of-cycle range 538 – 848

Registrations

7,673 → 8,802

end-of-cycle range 8,147 – 9,131

This month (Aug)

+35

466 → 501 expected by 31 Aug

Ensemble spread

538 – 848

pipeline 848 · seasonal 538 · momentum 541

The wide end-of-cycle band (538–848) is honest uncertainty, not noise: the three methods disagree on the Aug–Oct tail. The pipeline view (students already in the funnel) reads high at 848; the seasonal and momentum views read ~540. Expect the band to tighten as the tail actually declares.

The horizon ladder — today

Each step adds what last year delivered in that window, scaled by this year's current pace. This is the live projection your team reads.

Checkpoint	Admissions	+ from now	Registrations	+ from now
Now (12 Aug)	466	—	7,673	—
End of day	467	+1	7,701	+28
End of week (16 Aug)	477	+11	7,811	+138
End of month (31 Aug)	501	+35	8,417	+744
End of cycle	662	+196	8,802	+1,129

The one visual

The end-of-cycle call rose all season — because the season ran hot

Each point is what the engine expected the full-cycle admissions total to be, using only data it could have known on that date. It climbed from ~440 in April to ~660–700 by August as real admissions kept outrunning last year's pace. The shaded band is the engine's own low–high range; the dashed line is admissions actually booked to date.

Expected end-of-cycle admissions Engine low–high band Admissions booked to date

Every week's end-of-cycle call

The full record. "Booked" is admissions actually in on that date; the gap to "expected" is what the engine projected still to come.
As of	Booked	Expected adm	Adm range	Expected regs
20 Apr	17	441	143–642	2,941
27 Apr	20	418	172–586	3,534
4 May	28	425	161–640	2,947
11 May	43	533	178–884	3,051
18 May	56	400	191–554	3,104
25 May	66	358	221–467	3,139
1 Jun	81	392	247–512	3,374
8 Jun	111	448	295–593	3,789
15 Jun	136	453	375–519	4,421
22 Jun	184	469	434–510	4,982
29 Jun	224	490	449–540	5,473
6 Jul	260	531	484–594	5,997
13 Jul	304	599	528–691	6,808
20 Jul	340	610	497–767	7,805
27 Jul	392	661	525–851	8,494
3 Aug	433	697	538–931	9,502
12 Aug	466	662	538–848	8,802

Accuracy by horizon

Tight up close, cautious far out

WAPE = weighted absolute percent error (lower is better). Every row is scored against what actually happened. The tier reflects how much to trust that horizon.

Admissions — the number leadership tracks

Horizon	Scored	Avg miss	WAPE	Band hit-rate	Verdict
End of day	16	1.7	1.0%	—	near-exact
End of week	16	13.0	6.8%	—	reliable
End of month	15	13.6	6.4%	—	reliable
1 month ahead	11	37.5	25.0%	46%	directional
2 months ahead	6	50.5	28.9%	67%	directional
3 months ahead	2	61.0	33.0%	50%	rough

Registrations — strong near-term, weak month-ahead

Horizon	Scored	Avg miss	WAPE	Band hit-rate	Verdict
End of day	16	15	0.5%	—	near-exact
End of week	16	180	5.1%	—	reliable
End of month	15	531	13.8%	—	usable
1 month ahead	11	1,665	72.6%	9%	unreliable
2 months ahead	6	2,323	85.2%	0%	unreliable
3 months ahead	2	2,808	88.5%	0%	unreliable

Predicted vs actual

The month-end admissions call, week by week

Every Monday in July the engine predicted where admissions would land by 31 July. The actual was 418. The forecast stayed within a tight ±30 the whole month, and the day-of number was essentially exact all season.

July admissions — end-of-month prediction converging on the actual (418)
Forecast made	Days out	Predicted (31 Jul)	Actual	Error
6 Jul	25	399	418	−19
13 Jul	18	446	418	+28
20 Jul	11	396	418	−22
27 Jul	4	424	418	+6

Day-of accuracy is the standout. Across all 16 weeks, the engine's same-day admissions count missed by an average of under 2 students — the number that goes into a daily standup, essentially exact.

Same-day admissions — a near-perfect track

As of	Predicted today	Actual	Error
15 Jun	141	136	+5
29 Jun	225	224	+1
13 Jul	308	304	+4
20 Jul	345	340	+5
27 Jul	394	392	+2
3 Aug	434	433	+1

Where it misses — new registrations a month out

The one genuinely weak call. The engine expected new registrations to fall away sharply after each date; instead they held near ~3,000. This is why month-ahead registration volume is flagged unreliable — and why admissions, which lean on the existing pipeline instead, stay accurate anyway.

Registrations predicted to arrive in the month after each date, vs what actually arrived
As of	Predicted (month+1)	Actual	Error
1 Jun	427	3,171	−2,744
15 Jun	639	3,171	−2,532
29 Jun	999	3,171	−2,172
6 Jul	5,150	7,045	−1,895

How the engine works

Three independent methods, blended

No black box and no machine learning — the forecast is a deterministic ensemble of three views, so the same inputs always give the same number and every figure can be traced to its basis. Admissions accuracy comes from leaning hardest on the method that reads the existing funnel rather than guessing at new demand.

Method	What it reads	Today's EOC call
Pipeline	Students already in the funnel × last year's stage-to-stage conversion and timing	848
Seasonal	This year's pace projected along last year's month-by-month shape	538
Momentum	The last 28 days' run-rate carried onto the remaining curve	541
Blended	Pipeline-weighted ensemble; band = the spread across methods	662

Read this honestly

Where it's strong, and where to be careful

Trust the near term as a number. Today / this week / this month for admissions are all inside ~7% error, and the day-of count is near-exact. Plan against these.
The far horizon under-shot — systematically. Every 2–3 month-ahead admissions prediction came in low (e.g. an April call of 128 for what became 185). The model anchors on last year's seasonal shape, and this cycle accelerated past it. Treat long-range admissions as a floor, not a point estimate.
Registration volume a month+ out is not reliable (73–89% error). Last year's curve can't see this year's campaign calendar. Notably, admissions accuracy holds up anyway, because the pipeline method leans on students already in the funnel rather than on predicting new registrations.
The bands are honest up close, over-confident far out. Near-term ranges behaved; the month-ahead registration band caught the actual only 0–9% of the time — it was too narrow. The current wide end-of-cycle band (538–848) is the model correctly widening under genuine tail uncertainty.

Bottom line

How to use this forecast

Operationally, trust it. Today, this week, and this month's admissions are accurate to within ~7%, and the day-of count is near-exact. It is safe to plan staffing, seat-blocking, and weekly targets against these.
Strategically, read the direction and the band. The end-of-cycle call is 662 admissions (538–848). Given the season has out-run last year all year, the upper half of that band is the more likely story — treat 662 as a conservative center.
Don't quote a month-ahead registration number. That is the one figure the backtest says to withhold; use the admissions projection instead.
It improves as the cycle closes. Error shrinks sharply at short horizons, so the closer to a checkpoint, the more precise the call — most valuable in the final weeks when decisions are made.