Ablation report
Which architectural components actually carry MAS performance? We run the full MAS plus three lesioned variants — ADWIN drift detection removed, dynamic safety stock removed, and forecaster held at the simplest tier (MA only) — on the catastrophic scenario (10 seeds per variant) and measure how much each removal degrades stockout rate, total cost, MAPE, and the drift-event count.
Δ vs full columns are percent-relative to the full MAS row, so a positive Δ on stockout or cost means the lesion makes things worse. Near-zero Δ would suggest the removed component isn't pulling its weight; a large Δ is evidence the component is essential.
| Variant | N | Stockout | Δ vs full | Total cost | Δ vs full | MAPE | Global drift |
|---|
Reading the table: under the adversarial catastrophic scenario every individual ablation improves both stockout rate and total cost relative to the full MAS. The mechanism is documented in thesis §5.3 — the Holt–Winters forecaster captures the genuine transient volatility, inflating z·σ·√LT and driving over-provisioning. Bounded safety stock (capping the buffer at a multiple of mean lead-time demand) is the fix queued as future work.