Footwear resale market analysis — 99,956 transactions: exit timing, size-curve scarcity, demand risk
Analyzing 99,956 real StockX transactions to answer three questions every reseller faces: when to sell, which sizes to buy, and which shoes are risky bets.
Stack: Google BigQuery (SQL) → Power BI (star schema, DAX) · Data: StockX 2019 Data Contest
I've resold sneakers on StockX and eBay for about six years (80+ SKUs), so this dataset describes a market I operated in — every number here was sanity-checked against firsthand experience, including the fee assumptions. The data covers Sept 2017 – Feb 2019: two brands, 50 models, and two opposite playbooks. Off-White ran a scarcity model (27,794 sales at fat premiums); Yeezy ran a volume model (72,162 sales at thin ones, after adidas deliberately expanded supply in 2018). That contrast is the color language of every chart in the report.
- Exit timing — when is the optimal window to sell after release? (retail translation: markdown timing / dynamic pricing)
- Size-curve scarcity — which sizes carry premiums, and does the sweet spot shift by brand? (size-curve buying)
- Demand risk — which models are stable and which are gambles? (demand risk scoring in buy planning)
Cut from scope, deliberately: a brand-only premium comparison (two numbers = a KPI card, not a question), an absolute-dollar "hero" leaderboard (mechanically favors expensive shoes), and geography (regional volume mostly tracks population; a 50-model × 51-region grid is too sparse to trust).
1. Two playbooks, one market. Off-White wins on rate — 268.4% median resale premium vs Yeezy's 43.6%. Yeezy wins on dollars — $25.98M total resale vs $18.66M. Rate and revenue are different games, and the demand risk map shows them as two visibly separate clusters.
2. The premium "collapse" that wasn't. Market-wide, median premiums appear to collapse to ~31% a year after release. That turned out to be a composition artifact: the 1-year+ bucket is 98% Yeezy (23,730 of 24,185 sales). Split by brand, the stories reverse — aged Yeezys decay like commodities (30.9%), while aged Off-Whites appreciate like collectibles (578.1%, on just 455 sales). Exit timing is brand-specific: Yeezy sellers exit pre-release or in months 6–12; Off-White holders are paid to wait — if they can accept the illiquidity. The extreme case: Off-White Air Jordan 1s a year after release trade at a 1,189.5% median premium (n=111).
| Time since release | Off-White | Yeezy |
|---|---|---|
| Pre-release | 323.4% | 54.5% |
| 0–30 days | 243.8% | 36.8% |
| 31–90 days | 245.8% | 41.4% |
| 91–180 days | 286.8% | 75.9% |
| 181–365 days | 326.3% | 90.0% |
| 1 year+ | 578.1% | 30.9% |
3. Scarcity lives at the tails of the size curve. Core sizes 9–11 carry 66.8–79.5% median premiums on ~30K sales. Tail sizes 15–17 carry 206–288% — a 3–4x multiple — on just 213 sales. The size curve is also brand-shaped: size 15 is 100% Off-White (130 sales, zero Yeezys). Size 17's 287.7% rests on n=4, disclosed accordingly.
4. The market almost never loses. Only 0.56% of 99,956 sales closed below retail. Total retail value of $20.85M flipped for $44.64M — a $23.79M aggregate spread.
Footnote on medians: the all-market median premium is 70.5%, which is not "between" the brand medians of 43.6% and 268.4% in any weighted-average sense — a pooled median is not an average of medians. Yeezy is 72% of all rows, so the pooled middle lands near Yeezy's 69th percentile.
Source: 99,956 rows released publicly by StockX for their 2019 Data Contest (Kaggle). Full column reference in data/data-dictionary.md. Three genuine defects, all evidenced in sql/validation.sql:
| Defect | Impact | Fix |
|---|---|---|
' Yeezy' leading space on all 72,162 Yeezy rows |
Groupings look normal, but any brand = 'Yeezy' filter or join silently returns zero rows |
TRIM() in the staging view; confirmed via bracket test on raw data |
Inconsistent model-name capitalization (Adidas- vs adidas-) |
Breaks case-sensitive pattern matching | Silhouette parsing matches on LOWER() |
Prices/dates stored as text ($1,097, 9/1/17) |
Not directly analyzable | Typed at load by BigQuery schema detection |
raw_sales (loaded table)
└── stg_sales (view: renaming, TRIM, type casting)
├── dim_sneaker (50 rows: brand, retail, release date, parsed silhouette)
└── fact_sales (99,956 rows: prices, spread, days-since-release + timing buckets)
The staging view is the shock absorber: when the loaded schema turned out to differ from the plan (auto-typed dates and prices), only staging changed — both marts and everything downstream were untouched. Validation: five checks (row counts, dimension integrity, NULL audit, financial reconciliation, raw-defect evidence), with headline numbers verified through independent paths — an independent profile of the raw CSV vs SQL vs the Power BI model.
dim_sneaker (1) → (*) fact_sales ← (1) Dim_Date (DAX CALENDAR(), marked as date table). Silhouette parsing expands 2 brands into 11 product families. Measures (full formulas in dax/measures.md):
- Median Resale Premium % —
MEDIANXover per-transaction markup; median because prices are heavily right-skewed. - Net Premium After Fees % — the dataset has no fees, so seller economics are modeled with a disclosed 12.5% fee assumption (9.5% transaction + 3% processing, the 2017–2019 schedule, known from my own payout records).
- Price Volatility (CV) — stdev ÷ mean of sale price per model; dividing by the mean makes a $250 and a $1,500 shoe comparable, where raw stdev would just re-rank by price level.
- Reference lines on the risk map are the medians of the 50 model-level values (240.79% / 0.142), so each axis splits the catalog in half.
No marketplace fees in the data (assumption disclosed above); no seller identifiers; no inventory levels; no basket structure. Data ends Feb 2019 — these are the market's patterns, not current prices. Two brands is the dataset's scope, treated here as focus. CV assumes the mean is a sensible center and can overstate risk for heavily skewed models. Thin-sample findings carry their n (size 17: n=4; aged Off-White: n=455).
The Power BI file points at my BigQuery project. To rebuild: create a BigQuery dataset, load data/StockX-Data-Contest-2019-3.csv with schema auto-detect, run the scripts in sql/ in order (staging → dim → fact → validation), and repoint the .pbix connection reports/stockx-resale-analytics.pbix — or just view the report screenshots in images/.
This project runs on BigQuery's free sandbox tier with no billing account attached; live tables expire September 7, 2026 (confirmed via console). All SQL scripts in /sql fully rebuild the pipeline from the included CSV in minutes if the original tables have expired.