Index Methodology

How every number on this site is calculated, what it can and cannot tell you. Last rebuilt 2026-08-09 13:47 UTC.

Read this first. We publish three different kinds of number and they are not interchangeable. A measurement estimates how the whole market moved. A benchmark tracks a fixed basket of specific cards. A leaderboard is just a ranked list. Only the first two are indices.

Tier 1 — The Measurement (headline number)

The primary GPK market measure is a repeat-sales regression (RSR) in the Case–Shiller / Mei–Moses tradition. It is the only method here that measures the market rather than a basket, because it compares each card against itself.

The math

For every pair of sales of the same card (same spec — and grade is part of identity, so a PSA 10 and a PSA 9 are different cards), we take the log price change:

yₐ = log(price₁) − log(price₀)

Each pair contributes a row with −1 in the quarter it was bought and +1 in the quarter it was resold. Solving that system gives log index levels. Then, because Case & Shiller showed that price noise grows the longer a card is held, we run the standard three-stage correction:

Stage 1 (OLS) — fit the time dummies, keep the residuals.
Stage 2 (variance model) — regress squared residuals on the holding interval: e² ≈ A + B·Δt. Our fitted B is positive, which confirms the effect is real in this market, not just an assumption we imported.
Stage 3 (GLS) — refit weighting each pair by 1/√(A + B·Δt), so long-held pairs count less than tightly-spaced ones.

Base period 2017Q3 = 100. The same math produces the Vintage (1985–88) and Modern (2013+) sub-indices — they are segments of one model, not separate methodologies.

Current levels and the pair counts behind them

Segment	Latest quarter	Level	Pairs this quarter	Pairs all-time	Thin?
ALL	2026Q3	320.39	891	31,326	no
MODERN	2026Q3	276.66	229	4,191	no
VINTAGE	2026Q3	311.01	657	26,784	no

We publish pair counts because an index built on thin data must say so. Any quarter under 20 pairs is flagged thin on the chart rather than quietly smoothed.

Biases we accept, stated plainly

Repeat-sales selection bias. Only cards that sell twice enter the index. Cards that never trade again are invisible to it.
Survivorship. Cards that stop trading — often the weakest — drop out of later periods.
Grading events. We pair like-for-like within a spec, but a card cracked and regraded is a different spec and its history does not carry over.
Sale-price fidelity. We use all-in price where available. Fees, shipping quirks and private-sale terms we cannot see add noise.

Tier 2 — The Benchmarks (GPK25, Modern 25)

A benchmark tracks a fixed basket of named cards. It answers "how did these 25 do", not "how did the market do". Rules that make a benchmark auditable:

Equal-weighted means averaging constituent RETURNS, not summing basket value. If the label says equal-weighted, the math must be an average of returns — a value sum is price-weighted by another name.
A divisor is adjusted at every reconstitution so the level is continuous across composition changes. Swapping a card must never create a jump that looks like a market move.
Real sales only. No constituent is priced off an estimate. A constituent with no sale for 180 days is replaced at the next reconstitution, and the replacement is disclosed.
Concentration caps: at most 4 constituents per set and at most 1 per character. Adam Bomb holds one seat, not three.
Grade is part of identity. A PSA 10 and a PSA 9 of the same card are separate securities.
Staleness is published — the share of constituents with a sale in the trailing 90 days appears on the index page. A benchmark whose members have stopped trading is telling you less than it appears to.

Backtest disclosure. Where an index level predates the day we began computing it live, that span is reconstructed from historical sales, not observed in real time, and the chart labels it backtested. Reconstructed history is fitted with the benefit of hindsight and should be read more sceptically than the live span.

Tier 3 — The Leaderboard

The Top-25 Most Valuable list is a leaderboard: a ranked snapshot of the highest values right now. It has no level, no percentage since inception, and it is not an index. It changes when values change, which is exactly what a ranked list should do and exactly what an index must not do.

Limitations we will not paper over

GPK is a thin market. Roughly 20–25 sales are recorded per day across the whole category. Short windows are noisy; treat quarterly and annual movement as the signal.
Our sale data is PSA-verified by policy. Sales we cannot tie to a cert do not price anything.
Values shown on card pages may be estimates where sales are absent; those are labelled, and estimates never enter any index.
Population data comes from third-party census sources and can lag.

Questions or a correction? Every number here is reproducible from the published sales data. Data notes & corrections