Demand Forecasting
Method: weekly demand modelled as linear trend + 3-harmonic Fourier annual seasonality + promotion dummy, fitted by least squares per SKU across all four DCs. Chosen over a plain moving average because every SKU shows both a persistent trend and a stable 52-week cycle, and over ARIMA because the seasonal signal is deterministic and the promotion effect needs an explicit regressor. Holdout: last 8 weeks of history (weeks 97-104), scored per SKU x DC and rolled up on demand weights.
Seal Kit (SKU-005) — weekly demand, all DCs
104 weeks of history vs 12-week forecast · shaded band = holdout test period and forecast interval
Pattern detected
Hydraulic Components · ABC class A
- Trend
- +15.9% / yr
Growing base demand; a moving average would systematically under-forecast this SKU.
- Seasonality
- ±14.8% amplitude
Stable 52-week cycle captured with three Fourier harmonics; peaks repeat in both history years, so it is seasonality rather than noise.
- Promotion effect
- +58.1% lift
Estimated from 10 flagged promotion weeks and held out of the baseline. The forward forecast assumes no promotions; add the lift for any planned event.
- Forecast level
- 269.1 units / weekvs 248.6 historical avg
Handling the stockout-censored weeks (weeks 40-45, Asia Pacific DCs)
Trackline Manufacturing supply disruption affecting all five Undercarriage SKUs at Chennai and Jakarta
During weeks 40-45 the two Asia Pacific DCs were out of stock on Undercarriage parts. The recorded Units_Sold for those weeks is therefore not demand — it is the truncated quantity that happened to be on the shelf. Feeding it to a forecast model teaches the model a demand dip that never happened, and that dip flows straight into a lower safety stock for the least reliable supplier in the network — precisely the wrong direction.
Decision: correct the history rather than drop it. Dropping six consecutive weeks would punch a hole in the seasonal signal at the same point in the annual cycle. Instead each censored SKU x DC week is rebuilt from the same SKU's demand at the two unaffected DCs over the same weeks, normalised to each DC's own pre-disruption baseline (weeks 34-39) and re-scaled to the censored DC's baseline. Promotions and seasonality are shared across DCs while the outage is warehouse-specific, so the unaffected DCs carry the signal for what demand would have been. The regression is then fitted on the restored series.
A stricter alternative is a censored (Tobit-style) likelihood that models the stockout flag explicitly. At six weeks out of 104 the cross-DC index gives near-identical fitted parameters at a fraction of the complexity, so it was preferred; the original Units_Sold and the stockout flag are both retained so the choice can be revisited.
| SKU | DC | Observed | Restored | Uplift |
|---|---|---|---|---|
| SKU-011 | WH-AP1 | 123 | 327 | +166% |
| SKU-015 | WH-AP1 | 105 | 263 | +150% |
| SKU-011 | WH-AP2 | 55 | 156 | +184% |
| SKU-012 | WH-AP1 | 41 | 135 | +229% |
| SKU-015 | WH-AP2 | 67 | 136 | +103% |
| SKU-013 | WH-AP1 | 37 | 100 | +170% |
| SKU-013 | WH-AP2 | 18 | 57 | +217% |
| SKU-014 | WH-AP1 | 23 | 62 | +170% |
| SKU-012 | WH-AP2 | 25 | 56 | +124% |
| SKU-014 | WH-AP2 | 8 | 28 | +250% |
Forecast accuracy — 8-week holdout (weeks 97-104)
WMAPE weights errors by volume, so it is not distorted by low-volume weeks the way MAPE is.
| SKU | Product | ABC | Actual (units) | Forecast (units) | WMAPE | Bias | Naive WMAPE |
|---|---|---|---|---|---|---|---|
| SKU-005 | Seal Kit | A | 2,470 | 2,583 | 24.4% | +4.6% | 24.1% |
| SKU-009 | Air Filter | A | 2,525 | 2,528 | 26.2% | +0.1% | 31.9% |
| SKU-018 | Wiring Harness | A | 2,838 | 2,287 | 21.8% | -19.4% | 20.7% |
| SKU-011 | Track Chain | B | 1,483 | 1,458 | 19.3% | -1.7% | 20.1% |
| SKU-001 | Hydraulic Pump | C | 191 | 162 | 23.6% | -15.1% | 23% |
| Portfolio average | 23.1% | 24% | |||||