# SKU Rationalization Scoring Methodology

**Version:** 1.0  
**Data source:** Cinderhaven Postgres, 2023–2026 window  
**Portfolio:** 50 SKUs, 5 product lines, 6 retailers

---

## Overview

Every SKU in the portfolio is scored across five dimensions. Each dimension produces a score from 1 (worst) to 5 (best), calibrated against the actual distribution of that dimension across the portfolio. The scores are combined into a weighted composite and each SKU is assigned to one of four action buckets.

The framework answers: *which SKUs should be killed, fixed, maintained, or doubled down on — and what does each decision cost or save?*

---

## The Five Dimensions

### 1. Velocity
**What it measures:** Average units sold per store per week (USPW).  
**Direction:** Higher is better.  
**Data source:** `raw.scan_data`, aggregated per SKU across the full data window.

| Score | Threshold |
|-------|-----------|
| 5 | ≥ 13.51 USPW (p75) |
| 4 | ≥ 8.33 USPW (p50) |
| 3 | ≥ 4.36 USPW (p25) |
| 2 | ≥ 2.02 USPW (p10) |
| 1 | < 2.02 USPW |

### 2. Contribution Margin
**What it measures:** Loaded contribution margin as a percentage of gross revenue, after deducting COGS, trade spend, chargebacks, and allocated retailer deductions.  
**Direction:** Higher (less negative) is better. All Cinderhaven SKUs carry negative loaded margins after full cost loading; scoring is portfolio-relative.  
**Data source:** `public_intermediate.int_loaded_contribution_by_sku`

| Score | Threshold |
|-------|-----------|
| 5 | ≥ −4.37% (p75 — least negative) |
| 4 | ≥ −4.76% (p50) |
| 3 | ≥ −6.97% (p25) |
| 2 | ≥ −7.77% (p10) |
| 1 | < −7.77% |

**COGS formula (B2B):** `units_ordered × case_pack_qty × cogs_per_unit`  
(cases shipped × units per case × cost per unit)

### 3. Shelf Space Cost
**What it measures:** Annual cost of maintaining shelf presence, combining actual promotional spend with a $400/store/year overhead proxy for compliance, data, and resets.  
**Direction:** Lower is better (inverted scoring).  
**Data source:** `public_intermediate.int_shelf_space_cost_by_sku`

| Score | Threshold |
|-------|-----------|
| 5 | ≤ $40,136/year (p25 — lowest cost) |
| 4 | ≤ $74,765/year (p50) |
| 3 | ≤ $119,628/year (p75) |
| 2 | ≤ $137,925/year (p90) |
| 1 | > $137,925/year |

### 4. Production Complexity
**What it measures:** Ratio of landed cost per unit to MSRP (`landed_cost/msrp`). A lower ratio signals a simpler product relative to its price point — lower ingredient and manufacturing complexity.  
**Direction:** Lower is better (inverted scoring).  
**Data source:** `raw.sku_costs` joined to `raw.product_master`

| Score | Threshold |
|-------|-----------|
| 5 | ≤ 0.2475 (p25 — lowest ratio) |
| 4 | ≤ 0.2701 (p50) |
| 3 | ≤ 0.3012 (p75) |
| 2 | ≤ 0.3167 (p90) |
| 1 | > 0.3167 |

### 5. Cannibalization Risk
**What it measures:** Velocity penalty when sibling SKUs (same product line) are co-distributed in the same store. Expressed as the inverted velocity delta: `max(0, -(shared_uspw − solo_uspw) / solo_uspw)`. A value of 0 means no measurable cannibalization.  
**Direction:** Lower is better (inverted scoring).  
**Data source:** `public_intermediate.int_cannibalization_pairs`

**Methodology note:** This is a *proxy* (cross-sectional velocity comparison), not a rigorous difference-in-differences estimate. All 50 Cinderhaven SKUs were authorized within the same six-month window (January–June 2024), leaving insufficient pre-variant temporal depth for DiD. The proxy compares velocity in stores where a SKU is alone in its product line versus stores where sibling SKUs are also present. SKUs with fewer than 3 solo stores are treated as having no measurable signal (score 5). Because more than half the portfolio shows no cannibalization signal, the p50 threshold is 0.000 and no SKU is assigned a score of 4 — in practice scores fall at 5, 3, 2, or 1.

| Score | Threshold |
|-------|-----------|
| 5 | = 0.000 (no signal) |
| 4 | ≤ 0.000 (p50) |
| 3 | ≤ 0.0745 (p75) |
| 2 | ≤ 0.2054 (p90) |
| 1 | > 0.2054 |

---

## Threshold Calibration

All thresholds are derived from the actual percentile distribution of each dimension across the 50-SKU portfolio. They are not intuition-based.

To recalibrate against updated data:
```bash
flyctl proxy 5432:5432 -a cinderhaven-db   # in a separate terminal
python scripts/calibrate.py
```

This overwrites `src/scoring/constants.py` with new percentile values. Recalibrate whenever the underlying data window changes materially.

---

## Quadrant Assignment

Quadrant is determined by counting *red flags* — dimensions that score 2 or below — not by the weighted composite score. The composite score ranks SKUs within quadrants; it does not determine the quadrant.

| Red flag count | Action bucket |
|----------------|---------------|
| 2 or more | **Kill** — multiple structural failures; candidate for discontinuation |
| Exactly 1 | **Fix or Kill** — one critical weakness; address it or cut |
| 0, but min score < 4 | **Maintain** — no critical failures; hold and optimize |
| 0, and all scores ≥ 4 | **Double Down** — strong across all dimensions; invest and protect |

---

## Weighted Composite Score

The composite score (1–5) is computed as:

```
score = Σ(dimension_score × weight)
```

Default weights are equal (20% per dimension). The interactive demo tool at `app/index.html` allows weights to be adjusted per the client's priorities. Adjusting weights re-ranks the composite but does not change action bucket assignments.

---

## Limitations and Caveats

1. **Cannibalization is a proxy.** Cross-sectional comparison is directional, not causal. A positive cannibalization signal may reflect store-mix differences rather than true demand transfer. Rigorous DiD would require a longer pre-authorization window.

2. **All margins are negative.** Cinderhaven's fully loaded margins are negative across the portfolio after accounting for COGS, trade spend, chargebacks, and deductions. The scoring is portfolio-relative: score 5 means "least negative," not profitable.

3. **Thresholds are portfolio-relative.** A SKU that scores 5 on velocity in this portfolio might score 3 in a different brand's portfolio. Thresholds should be recalibrated for each new client engagement.

4. **The $400/store/year overhead proxy** for shelf space is an estimate. It covers data compliance, reset labor, and retailer relationship overhead. Adjust this constant in `int_shelf_space_cost_by_sku.sql` if the client has actuals.

5. **Complexity proxy does not capture R&D or packaging complexity.** It uses `landed_cost/MSRP` as a signal only. A more rigorous version would require a bill-of-materials and production time data.