Forecast Accuracy (Backtested)

Enterprise scope · 25,693 item-DCs scored · rolling-origin backtest over the last 12 months of their own history — forecast accuracy measured today, no waiting for a live track record

Today's Implicit ForecastiA min/max policy implicitly forecasts 'next month = recent average'. This is that forecast's weighted absolute percentage error, backtested against what actually happened. Lower is better.

45%

WAPE of the flat average min/max assumes

Per-Item Champion MixiEach item-DC scored on the method that backtests best on ITS history — seasonal methods on seasonal items, Croston/SBA on intermittent ones. This is what a deployed forecasting layer would do. (Champion selection and scoring share the holdout, so treat this as an upper bound on the improvement.)

39%

each item-DC on its own best method · best single method 46% (Croston/SBA)

Item-DCs ScorediRolling-origin backtest coverage: item-DC combos with at least 12 training months plus a scoreable holdout. The rest (very new or no-demand items) can't be scored honestly and aren't.

25,693

every combo with ≥15 months of history

WAPE by demand class × method

Darker = worse. The pattern to notice: no single method wins everywhere — class-aware method selection is the unlock. Values are exact aggregate WAPE, not averaged ratios.

Method	Smooth	Erratic	Intermittent	Lumpy	Overall
Flat average (your min/max's implicit forecast)	33%	76%	57%	125%	45%
Seasonal naive (same month last year)	41%	79%	72%	148%	54%
Moving average × seasonal index	37%	72%	57%	129%	48%
Croston/SBA (intermittent-demand estimator)	32%	77%	60%	139%	46%

Largest forecast misses (champion method)

Item-DCs with the largest absolute unit error over the holdout — where better forecasting (or a demand conversation with the customer) matters most.

SKU	Description	DC	Best method	WAPE	Bias
813908	SW2 10 04 WHT	110	Croston/SBA	34%	+2%
813909	SW2 10 06 WHT	110	Croston/SBA	41%	+16%
464000574	COIL 029X3/4 GA/G90	3510	Croston/SBA	77%	+2%
703908	682M 10 04 W	110	Flat average	18%	-16%
H880	6" RD X 2" LONG WT PVC SPLIT PIPE FOR QS6	109	Flat average	110%	+14%
010770	800 10 04 BRN	110	Seasonal naive	36%	+5%
P703909	682MPBS 10 06 W	127	Flat average	86152%	+86152%
016108	4REA90	110	Croston/SBA	15%	-9%
57007	INDIVIDUAL UNIV INSIDE/OUTSIDE CORNER WHITE PVC	109	Croston/SBA	919%	+919%
813920	SW3 10 06 WHT	110	Croston/SBA	33%	-20%
02050804CW	140MC 08 04 CW	127	Moving average × seasonal index	45%	-6%
3000250	HANDLE RIGHT 900	127	Seasonal naive	100%	+50%
P703908	682MPBS 10 04 W	127	Croston/SBA	3293%	+3293%
810014431	Sidewall Termination Kit- 21" (533mm)	3540	Seasonal naive	119%	-72%
703909	682M 10 06 W	110	Croston/SBA	31%	+18%

Method: for each of the last 12 months, train only on history before that month, forecast 1–3 months ahead with each method, and score against actuals. WAPE = Σ|forecast−actual| / Σ|actual| (MAPE explodes on intermittent items, which dominate this catalog). Bias > 0 = over-forecasting. Benchmark: 20–35% WAPE is strong for make-to-stock hardgoods; their consultant's in-room benchmark was “80%+ accuracy would be world class” (= 20% error).