Trading-Card Sales Tapes — Foundations

How a card sales tape is shaped, where it is reliably broken, and which analyses on it are wrong in ways that look right.

This skill is written to be portable. The structure, defects and traps below recur across card sales tapes because they come from how these markets and pipelines work, not from one vendor. Concrete numbers come from a reference implementation and are marked as such:

Ref impl — measured against a Pokémon TCG lakehouse (~9.4M sales, Feb–Aug 2026, DuckDB over Parquet). Connection details, table names and environment quirks: references/reference-implementation.md.

Treat ref-impl magnitudes as order-of-magnitude priors and worked examples, not constants. Re-measure them on your tape — every section tells you how.

What a card sales tape is

It is not a price series. It is a transaction log in which the same card trades simultaneously in several observable markets — a marketplace's raw listings, an auction site's raw sales, graded copies at each grade, off-condition copies — and those markets do not move at the same time.

Nearly all exploitable signal lives in the relationships between those markets, not in the price history of any one. Card-level price momentum is close to worthless at short horizons; the cross-venue basis is typically the strongest single factor.

The practical consequence: work at card × venue × period granularity, with periods of 14 days or a month. Collapsing to a single blended monthly median per card throws away almost all the information and leaves you with a few dozen effective observations.

The five questions to answer before your first analysis

Any card tape you pick up. In this order:

1. Which venues are actually usable? Identity resolution rate per venue, and trade_type composition per venue. Most venues fail one or both.
2. How is condition/grade encoded, and is there more than one encoding? There usually is. Pick the complete one, not the well-named one.
3. Where does each venue's coverage start and stop? Per-venue distinct-days-per-period against the calendar.
4. How sparse is the panel? Fraction of card×period cells clearing your minimum sale count — broken out by era.
5. What is the measurement noise ceiling? Split-half reliability. This bounds every result you will ever get and usually matters more than model choice.

scripts/profile.sql runs all five. Do not skip to modeling.

Seven things that will burn you

1. There is probably no canonical grade column — find the complete one. Card pipelines accrete condition encodings: a marketplace's own vocabulary, a parsed-from-title vocabulary, a normalized slug vocabulary. They coexist, disagree, and neither dominates. Check both row count and price agreement before choosing. Agreement on price plus disagreement on count means one column is simply more complete — take it. Never fold in an "unknown condition" bucket: it is not "probably NM", it is a different, usually pricier population of unparsed listings.

Ref impl: two encodings; the slug one returned 2.36× more rows at a median price ratio of 1.000 across 1,826 cards. 678 distinct values across the two columns, 109,314 rows where they contradict, 893 rows on impossible grades (PSA 172, CGC 2006). Filed upstream as lake-of-rage#1580.

2. Assume a silent writer regression is in flight. A new ingest path lands, emits its own vocabulary for one or two columns, and every downstream consumer's filter silently drops its rows. This is invisible by construction — the rows are present and correct, just spelled wrong — so it shows up as an unexplained volume decline, which reads as a demand collapse. Profile distinct values of every filter column by month. A vocabulary that gains a new value mid-series is a regression until proven otherwise.

Ref impl: since a specific date, a new writer emitted a fourth vocabulary for both the grade and identity-confidence columns. 360,669 fully-identified rows (card key populated on 100% of them) were invisible to the canonical filter, and the share was still growing. Same window, same venue: old encoding 100% resolved, new encoding 0.0%.

3. Drop the newest period, always. The most recent bar is still filling. A signal built on it reads ingest lag as collapsing demand. This is universal and free to fix.

Ref impl: final-bar volume fell 66% — part lag, part the §2 regression, which alone stripped 43% of that period's rows on one venue.

4. Venue coverage windows differ, and a short one is usually expected. Venues get added to a pipeline at different times and backfill to different depths. Do not read a coverage gap as a market event, and do not file it as an outage — it is the shape of the data. It only determines where your panel can start: begin at the first period where every venue you need is adequately covered.

Ref impl: the second venue's history is shorter and has a 37-day interior gap. One period before the start bound had 48 sales on one venue against 48,836 on the other — a "venue gap" computed from that is noise.

5. Most "venues" are not secondary markets. A tape that aggregates on-chain and marketplace sources will carry primary issuance, buybacks, burns, internal transfers and vault movements alongside genuine arms-length resale. Filter on trade type explicitly. Headline row counts per venue routinely overstate tradeable volume by 5–10×.

Ref impl: one venue was 98.8% primary/buyback, another mostly burn/transfer, a third had zero secondary rows. Genuine secondary across all unresolved venues was ~148k, not the ~900k the raw counts suggested.

6. Lot sales stamp the lot price on every card. Multi-card transactions frequently write the whole lot's price onto each row. Guard with count(*) OVER (PARTITION BY <transaction key>) = 1, or reconstruct per-card value. Check this even if it is currently confined to venues you exclude — it goes live the day one of them is resolved.

Ref impl: in 4–10-card transactions, 86% shared one identical price — median $48/row against $5.73 for genuine single-card sales.

7. Never compare venues unconditionally. Venues differ in price formation, not just level: auction-style venues cluster on psychological ask points, bulk marketplaces are continuous and dominated by sub-dollar volume. A pooled cross-venue comparison measures product mix, not basis. Only compare per card, with a minimum sale count on both sides.

Ref impl: same era, same condition, median price by venue: $0.19 vs $10.00. All of that 50× is mix.

Two more rules for aggregation

• Compute returns within a venue, then volume-weight across. Venue mix moves violently as pipelines add and lose sources. Pooled medians fabricate large moves that are pure composition.
• Report within-set IC, not just total IC. A large share of forward variance is set-level. A factor with high total IC and near-zero within-set IC is a set-beta proxy — it cannot choose between two cards in the same set, which is the only decision that matters when buying one card.

Ref impl: ~34% of forward variance was set-level; one factor scored total IC 0.491 / within-set 0.034.

Sparsity is the binding constraint — reconstruct across conditions

Most of a card catalogue never trades. Expect the majority of catalogued cards to have zero sales in any given quarter. Every method question reduces to "does this buy me more observations per card-period?"

And rawness declines with print age. Older cards mostly did not survive in near-mint, so their tape is dominated by played-condition and graded copies. This is a property of physical card survival, so it should hold on any long-lived card line.

Ref impl — raw-NM share of sales by era: newest 64.8% · recent 64.2% · middle-modern 46.5% · mid-era 30.3% · vintage 29.4%. Mid-era tracks vintage, not modern. The same grouping appears independently in the graded premium (vintage 7.5× / mid 7.7× vs modern 1.5–1.8×) and the played-condition discount (0.766 / 0.728 vs 0.895 / 0.924) — three unrelated measurements, same split.

The panel consequence — fraction of card×period cells clearing a minimum sale count:

A best-condition-only panel is not viable on older eras. The fix is to convert every raw condition to a best-condition-equivalent price using a per-era factor, then pool.

Condition legs invert cleanly; graded legs do not. Condition ratios are tight enough for a multiplier; graded premiums vary far too much card-to-card. Use grades as a directional signal, never as a price level.

Ref impl: interquartile spread 1.50–1.95 for played conditions vs 3.49–4.25 for graded. Pooling raised split-half reliability in every era (mid 0.897 → 0.967) while mid-era usable cells went 3,299 → 17,584 (5.3×).

Factor tables, the recipe, and its validation: references/sparsity-and-eras.md.

The anchor model

In a multi-venue card market, one venue is the price anchor and the others error-correct toward it. With gap = ln(follower_price / anchor_price) per card, regress each venue's next-period move on the gap. The anchor's coefficient is ~zero (it is exogenous); the follower's is negative (it closes the gap).

This turns forecasting into convergence. You are not predicting what a card will be worth. You are measuring a published anchor, measuring how far a slower venue sits from it, and collecting the gap.

The wider implication: every unobserved venue — local shops, shows, social marketplaces — prices off the same public anchor and updates more slowly than the observed follower. So the follower's measured half-life is a lower bound on how stale in-person pricing is. That is the mechanism behind in-person arbitrage.

The anchor migrates with price band. Expect the bulk marketplace to anchor cheap cards and the liquid auction venue to anchor chase cards, with the crossover somewhere in the middle. Always fit this per band — a pooled correction coefficient is a cross-sectional average that matches no individual band.

Ref impl — follower correlation with gap −0.384 vs anchor −0.002; pooled gap decay ρ ≈ 0.78 per 14 days.

Price band	Follower vol share	β_follower	β_anchor	Anchor's share of correction	Gap half-life
<$10	75.2%	−0.282	0.013	4%	28 d
$10–25	79.0%	−0.308	0.046	13%	22 d
$25–50	79.8%	−0.215	0.050	19%	32 d
$50–100	84.2%	−0.158	0.061	28%	40 d
$100–250	86.7%	−0.478	0.154	24%	10 d
$250+	91.4%	−0.390	0.234	38%	10 d

Under $25 the bulk marketplace is truth. Above $100 the auction venue carries 87–91% of volume and the marketplace starts following it. Mid-band converges slowest (40-day half-life — the widest window to act in); the top band closes in ~10 days.

Demean the gap per card. The universe mean gap is nonzero — one venue runs structurally richer — so the raw sign is not the signal. Deviation from a card's own baseline is.

Ref impl: universe mean gap +0.16 logs, i.e. the follower ran ~17% richer.

Factor table

Measured on the ref impl at a 14-day horizon. The skip column is the honest one (see trap 1). Treat the ranking as a prior, the magnitudes as instance-specific.

Note the structural pattern, which should transfer: the factors that survive are the ones built from disjoint transaction sets — graded sales are different transactions from raw sales, played sales are different from best-condition. Signals built from the same transactions as the target are exposed to bounce.

Card momentum was negative at 14d and +0.20 at 60d — short-horizon reversal, long-horizon continuation. Do not mix horizons in one model.

Equal-weight composite (venue gap + volume growth + p10 change, percentile ranks, no fitting): IC 0.283 next-period, 0.167 skip-tested, within-set 0.225. Decile 1 → 10 spread −37.8% to +5.0% forward, monotone. The short side is much stronger than the long side — this is primarily an avoid-list. That asymmetry is worth checking for on any tape; overpriced cards are easier to identify than underpriced ones.

Dollar economics

For zero-transaction-cost (in-person) trading the objective is price × E[%move], not E[%move]. Cheap cards can have excellent percentage signal and no tradeable spread.

Ref impl — top-decile composite, forward 14 days:

Band	$ per card	% clearing $2+
<$5	$0.18	3%
$5–10	$0.37	1%
$10–20	$1.27	25%
$20–50	$2.22	43%
$50+	$3.37	29%

A price cap is usually the binding constraint on dollar spread, not signal quality.

And the threshold is channel-specific. $2 is a real spread in person at ~0% fees and nothing at all on a venue taking ~13%. Always state the channel a dollar figure assumes.

If you are here to build a model

Six properties of this asset class break assumptions that are so safe elsewhere they go unstated. Full treatment and evidence in references/asset-class-properties.md.

• It is a point process, not a time series. Every price series here is a grid you imposed on event data. Arrival intensity is a signal in its own right — often better than price.
• Observation is endogenous. Cards trade because someone wanted them. Dropout varies with both the signal and the price: in the reference implementation the top return quintile is 6pp less likely to trade again than a flat card, and $50+ cards drop out at 81.5% vs 93.0% for sub-$2. Your training set is selected on the outcome, and it loses exactly the expensive names with tradeable spread.
• Cold start is the normal case. New sets arrive continuously with zero history — 14.3% of the ending universe was new over three months. This is why hierarchical models earn their keep.
• Most of the observed return variance is measurement error. Reliability ~0.70 means the ceiling on correlation with true momentum is √0.70 ≈ 0.84. Evaluate against that, not against 1.0.
• Weighting can flip the sign. Same universe, same window: equal-weight −7.65%, dollar-weight +3.18%, median 0.00%. State which you chose.
• You cannot short. Long-short decile spreads report a number you cannot monetize. Report the long leg separately.

And you cannot back-fill history. Venues purge sold data on ~90-day horizons, so deep history exists only where someone was already capturing it. Broad early capture beats every model choice available.

Answer the person, not the panel

Nobody buying cards has a Sharpe ratio. Four personas ask, and they optimize different — sometimes opposite — things:

• Collector — buys to keep. Wants a fairness check on today's ask, not a return forecast. Reprint risk is their real enemy.
• Flipper — wants a buy-under and a sell-over price in dollars, plus how long the exit takes. A percentage on a $3 card is not an answer.
• Vendor — judged on turns, not appreciation. Ranking by expected return without liquidity hands them a shelf of unsellable stock. The avoid-list is worth more to them than the buy-list.
• Grader — wants EV of submission. We can supply the graded price distribution; P(grade) depends on the physical card and analysis must defer there.

Three things the hobby knows that the tape cannot see — treat as veto inputs, not features: reprint risk (modern supply is elastic; vintage is fixed), rotation and playability (a published, exogenous calendar), and condition in hand (centering, whitening and print lines appear in no column).

The most-asked question is the one the data cannot support. "Is this a good long-term hold?" over a six-period window is not answerable. Decline it rather than extrapolating a 14-day coefficient.

Personas, the lingo→column translation table, market-structure facts and the folk-claims register: references/personas-and-decisions.md and references/domain-knowledge.md.

References

• references/tape-anatomy.md — what a card sales tape contains, the columns that matter, and what to profile first.
• references/data-quality.md — the defect taxonomy, ordered by damage, each with a diagnostic.
• references/sparsity-and-eras.md — era structure, panel fill rates, the condition/grade ladder, and the cross-grade reconstruction recipe.
• references/methodology.md — the analytical traps. Every one has produced a wrong-but-plausible result.
• references/asset-class-properties.md — for ML/quant work: what card markets are structurally, and which standard modeling assumptions they violate.
• references/personas-and-decisions.md — collector / flipper / vendor / grader: what each is optimizing, what output is useful, and how their question translates into ours.
• references/domain-knowledge.md — the hobby's vocabulary mapped to columns, market-structure facts absent from every tape, and the folk-claims register (confirmed / retracted / untested).
• references/findings.md — dated result log from the reference implementation, plus superseded claims.
• references/reference-implementation.md — the concrete instance: connection, tables, columns, environment quirks.
• scripts/profile.sql — the five opening questions, as one script.
• scripts/panel.sql — canonical panel builder with guards.
• scripts/q.sh — remote query runner for the reference implementation.

Known open problems

Generic first, then instance-specific.

• One regime is the binding limitation on every coefficient here. All magnitudes were fit in a single up-then-cooling episode with no observed drawdown. Extending identity resolution backwards to reach a second regime is worth more than any modeling improvement.
• Ask-side data is the only true leading indicator. Sales say what cleared; listings say what is about to. Time-to-sale beats price as a demand signal. If your tape has a listings table, keep it fresh.
• Auction closes are arguably the cleanest clearing prices in any card tape, and a trade_type='secondary' filter usually excludes them. That exclusion is a choice, not a given.
• Exogenous drivers are absent from every column — reprints, rotation, tournament meta, a popular video. They are a large share of what actually moves prices. That is an information problem, not a modeling one.
• Population data and sealed product are usually missing or empty; check rather than assume.

Ref impl specifics: ~810k pre-2026 rows sit at <1% identity resolution; the main auction venue is only 43.9% resolved and the unresolved half is probably non-random (messier titles, bundles), which is where mispricing concentrates; the listings table is stale; the PSA population table is empty (PSA prices are not — 1.08M transactions); no sealed-product table exists. Breadth (% of a set's cards up) tested as coincident, not leading — do not build an early-warning indicator on it.