Fragrance Collapse Analysis

The −50% floor protects against underforecasting collapsing fragrances by applying a minimum. But there is no analogous upper bound. This asymmetric constraint is a systematic bias generator: it prevents one class of error at the cost of introducing the opposite class.

Backtest case — Coastal Tide (Y4): Actual decline was only −17%, but the −50% cap was applied, causing a 38% underforecast. The cap forces a pessimistic forecast on fragrances that don't deserve it.

This cannot be fixed by tuning the floor value. It requires either removing the floor entirely and letting the pooled decay curves carry the distribution, or implementing a symmetric upper bound so the constraint is directionally neutral.

Fragrance-level MAPE at n=5 is not a stable estimate. With only 5 fragrances, each contributes 20% of the metric — removing any single fragrance shifts the headline number materially. MAPE is also sensitive to small-denominator effects: fragrances with small actuals inflate percentage error regardless of forecast quality.

The headline fragrance-level MAPE must always be reported with this instability context — not as a precise model performance number. sMAPE (symmetric MAPE = 2 × |forecast − actual| / (forecast + actual)) is less sensitive to these effects and should be the primary reported metric.

The deeper issue: MAPE measures forecast accuracy, but inventory decisions are profit/loss decisions under uncertainty. The correct instrument is Newsvendor — it optimizes stocking quantities given a cost-of-underage vs cost-of-overage ratio, which is the actual business decision.

Confidence interval coverage in the backtest was far below nominal. CIs built on decay rate uncertainty alone showed only 1 of 5 actuals inside the 70% CI (expected: 3–4) and 3 of 5 inside the 90% CI (expected: 4–5). Coastal Tide's actual fell above the 90% upper bound; Pomelo's fell below the 90% lower bound — complete misses at the widest interval level.

Structural cause: Layer 3 growth rate uncertainty accounts for 63% of actual forecast error but is entirely absent from CI construction. Intervals will remain underconfident until growth uncertainty is added.

Monte Carlo path fixes this by drawing jointly from the decay distribution and the growth rate distribution — producing intervals that reflect the true uncertainty structure of the model.

The backtest does not test the pooled decay curve approach. In backtest mode (train on 2024 only), there are zero year-over-year transitions — so the pooled curve is empty and the engine falls back to the −50% cap for all returning fragrances.

The backtest is answering: "what happens if you apply a −50% decay cap to everything at 28.5% growth?" It is not a test of the pooled approach at all.

Any claim that this backtest validates the model design requires this disclosure. The pooled decay curves — the core innovation — have never been tested out-of-sample. Historical depth (more years of YoY data) is the prerequisite for a real test.

The 28.5% growth target is accepted as the planning input. What it currently lacks is a documented validation process — a set of leading indicators that could confirm or refine the target before each production run.

Given the engine's 239 unit/percentage-point sensitivity to the growth rate input, establishing a repeatable validation process is the highest-leverage reliability improvement available without any additional engineering. A 1pp error in the growth target propagates to 239 units of misallocation across the fragrance mix.

The fragrance collapse signals identified in this analysis — pre-season velocity cliffs, within-season trajectory divergence, low repeat retention — are exactly the leading indicators that could confirm or invalidate the growth target each cycle.