What Predicts a Pokemon Card's Return?

Six factors were run against the forward return of 41,359 cards over the twelve days to 2026-09-22, and almost none of them work. A card's trailing move predicts its next move with a rank correlation of -0.0016, which is another way of saying not at all. The median forward return is exactly 0.00% in every trailing-return bucket and every price bucket, and the reason matters more than the result: 21,676 of the 41,359 cards, 52.41%, carry a forward return of exactly zero, so for most of the catalogue a twelve-day window cannot resolve a move at all. That is absence of evidence rather than evidence of absence. The one factor with a measurable loading is set age, at 0.1193, and it is not monotone. This study is the superset of two earlier momentum pieces on this site, one of which we retracted, and it explains why the answer keeps changing.

We hold two price tapes and they have opposite shapes. The broad tape is a daily dump of the whole priced catalogue. It covers 52,656 products but only runs from 2026-08-30 to 2026-09-22, so it is wide and short. The deep tape is a per-card weekly series going back to 2025-01-27, but only 2,766 cards reach that first week, which is 13.92% of the 19,876 products on it. A card earns a place in the deep pool by having been continuously quoted and continuously interesting for the better part of two years. That is selection on the outcome, and it is the single largest reason published backtests in this hobby disagree with each other.

So every factor below is run twice. Panel A is the broad tape: a trailing window of 2026-08-30 to 2026-09-10 and a forward window of 2026-09-10 to 2026-09-22, eleven days and twelve days. Panel B is the deep tape: trailing 90 days against forward 90 days, stepped every 28 days, giving 40,823 point-in-time observations across those 2,766 cards. Neither panel is the truth. Reading them side by side is the closest we get.

Read the median column first, because it is the whole study. It is 0.00% in all eight rows. Whatever a card did over the previous eleven days, the typical card's quote was unchanged twelve days later. The mean column does move, and it moves against momentum: the best mean sits in the worst trailing bucket and the three strongest trailing buckets are the only negative ones. But look at what carries those means. In the down-more-than-20% row, 64.46% of cards did not move at all over the forward window and the win rate is 25.40%, so a +11.41% mean is being produced by a small number of quotes snapping back after a bad print. That is a repricing artifact, not a tradeable edge, and The Frozen Tape already measured the mechanism: 37.5% of week-apart quotes on this feed are identical to the cent.

Panel B looks like a different asset class. Every bucket is positive, medians run from 6.25% to 17.10%, and win rates sit between 59.44% and 74.02%. The rank correlation between trailing and forward return here is -0.0409, a faint reversal tilt rather than momentum. The difference between the two panels is not the horizon and it is not the method. It is which cards are in the room. The deep pool's median price is $4.44 against $1.77 for everything else on the same days, and the pool was assembled by requiring two years of continuous quoting, which is a filter on exactly the thing a return study is trying to measure.

Three of those four round to nothing. A Spearman of -0.0016 on 41,359 observations is not a weak effect, it is the absence of one, and a 0.0180 loading on price level is the same. The honest summary of the price factor is in the buckets rather than the correlation: mean forward returns fall from 3.63% under 25 cents to -0.39% above $500, but the median is 0.00% at every level and the cheap end's mean is manufactured by cards whose quotes tick between one and three cents.

Set age is the only factor here with a rank correlation worth printing: 0.1193 across 20,828 English cards whose set carries a recorded release date, measured over the twelve days to 2026-09-22. It is also the factor most likely to be misread, so three things need saying at once. First, the unit here is the CARD and the window is twelve days. Second, the relationship is not monotone: it rises to a plateau across the 10-to-20-year range, 4.44% and 4.20% in the two bands there, and then falls away to 0.87% and 0.13% for sets older than twenty years. The sentence is not that older is better; it is that a middle-aged set outran both the new ones and the genuinely vintage ones over these twelve days. Third, the effect is concentrated where quotes are thin. Inside the under-$1 band the correlation is 0.0478; inside the $25-to-$100 band it is 0.2414.

Rarity does not separate forward returns in any usable way. Across tiers with 200 or more cards the medians are 0.00% everywhere except a handful of modern parallel tiers, and those move in both directions: Double rare reads -7.22% at the median on 307 cards, Illustration rare -1.50% on 482, and Shiny rare +0.65% on 310. The largest of those is negative, which is worth saying plainly rather than leading with the two smaller figures. Language does not separate them either: English cards averaged 1.86% and Japanese 1.94% over the forward window. What language changes is not the return but the quoting. 61.79% of Japanese cards did not move at all against 43.53% of English ones. For what exists in how many printings and rarities, see the catalogue census rather than re-deriving it here.

That table is the closest thing to a finding in the negative results, and its last row is the only bucket in this study whose median is not 0.00%. Cards whose quote was rewritten six to eight times in eleven days are the only group where most of the population is actually being marked, 8.33% unchanged against 76.44% at the other end, and that group's median forward return is -0.96%. A median moves once the quotes do. 14,840 cards of 41,359 had a quote that was not rewritten once in eleven days, and 76.44% of them were still unchanged twelve days later. Nothing about a card's price predicts its return nearly as well as the fact that its price is not really a price. Whether a given quote can be trusted at all is measured separately in our work on price reliability and sample size; here it is an input, and the input says most of the catalogue is not being marked.

Two panels, both as of 2026-09-22. PANEL A: the `price_history` table, a daily dump of the priced catalogue, days 20695 to 20718 inclusive (2026-08-30 to 2026-09-22), 1,150,986 rows over 52,656 TCGplayer products, of which 44,711 join a `cards` row on `tcg_id`. A card enters Panel A only if it carries a price on all three of days 20695, 20706 and 20718, giving n = 41,359. The trailing signal is day 20695 to day 20706 and uses no data after 20706; the forward return is day 20706 to day 20718. PANEL B: the per-card weekly `ppthist:<tcgId>` series in `api_cache`, 19,876 of which join the catalogue. A card enters Panel B only if its series reaches 2025-01-27 and carries at least 40 points, giving 2,766 cards. Evaluation points sit on a single 28-day grid anchored on 2025-01-27, the tape's first recorded week, and an evaluation point is used only when the whole 180-day window fits inside that card's own tape, so no leg is resolved past either end of it. At each of the three required dates the price used is the LAST recorded point at or before that date, accepted only when it falls within 10 days of it. That convention gives 40,823 observations and a trailing-to-forward rank correlation of -0.0409. Every part of it matters and none of it is obvious, so the alternative is published beside it: resolving to the NEAREST point within 10 days either side and letting a window overhang the tape gives 43,821 observations and -0.0411. The two readings agree to the third decimal; the observation count does not, and a Panel B figure quoted without its grid rule is not reproducible. All prices on both panels are TCGplayer market prices, which are ASKS, not sold medians; no graded price appears in any return here. Every panel is keyed on the TCGplayer product id and carries exactly one row per product. 51,719 catalogue rows carry a TCGplayer product id, and they resolve to 50,507 distinct ids: 1,212 ids are claimed by two rows each, 2,424 rows in all. Those pairs are not a Japanese print against its English twin, which is what we assumed before counting them. Every one of the 1,212 is a same-language pair, 972 Japanese and 240 English, and they are overwhelmingly one product indexed under two catalogue spellings of the same set, as the XY trainer kits are. Each id collapses to a single row, preferring an English row and then the lexically first catalogue id, so no product enters a cross-section, a factor sort or a census twice. The tie-break is stated rather than assumed away, because it decides which of the two spellings a shared product is filed under. Rank correlations are Spearman, computed with average ranks on ties. Medians use the house convention, the upper-middle element on even samples. Every return bucket in this study is closed at its upper edge, so a card up exactly 25.00% is in the 10-to-25 band and not the 25-to-50 one; 88 cards sit on the first of those edges and 69 on the second, which is enough to move three printed row counts if the rule is left unstated. Set age joins `cards.set_name` to the 201 cached set-metadata rows carrying a release date. That join took four attempts and the first three were wrong, which is worth stating because the failure mode is invisible in the output. The first matcher reached 31.61% of the panel; its residual turned out to be a bug, not missing data, because the metadata stores provider-prefixed names (`sm cosmic eclipse`) against a catalogue that stores plain ones (`Cosmic Eclipse`). The second reached 86.86% of English cards, and its residual was ALSO mostly a bug: 40 of those 60 unmatched names, covering 2,358 cards, had a release date sitting in the metadata that the matcher could not reach, because `&` folded to a space rather than to `and`, because ` base set` was appended but never stripped, and because promo sets and series abbreviations are spelled differently on the two sides. The third reached 98.24% and was wrong in the other direction, which is the more dangerous one: it dated 206 cards to sets they do not belong to, and a false match is invisible where a miss is not. Its prefix stripper removed the first word of any catalogue name whether or not a provider prefix was there, so `Expedition Base Set` matched the metadata row `Base Set` and 165 cards were dated 1999-01-09 instead of Expedition's 2002-09-15. Its token rules dropped the words `base` and `set`, which left the metadata row `Base Set 2` carrying the single token `2`, and that token is a subset of both `POP Series 2` and `EX Trainer Kit 2: Plusle & Minun`, so 41 more cards were dated to February 2000. Both of those sets do have a metadata row and neither row has a release date, so both belong in the residual rather than in the panel. The matcher used here folds accents, apostrophes, gender signs, case and punctuation, maps `&` to `and`, normalises promo vocabulary, expands a listed closed set of series abbreviations, strips a provider prefix only where a colon or a dash marks one, refuses a token-subset match on a single numeric token, and then resolves exact, unique-token-subset, and unique-token-superset in that order. Two tie-breaks apply when a bare catalogue name matches several metadata rows: a base-set row beats a promo or energies row, and a row that adds nothing to the query but a series token wins outright, which is how the catalogue's `Dragon` resolves to `EX Dragon` and its 2003 release rather than to Dragon Vault or Dragon Majesty. English coverage is 98.06%, 20,828 of 21,240. The remaining 20 names covering 412 cards were checked one at a time against the metadata rather than assumed: 404 of those cards belong to sets carrying no recorded release date, which is POP Series 1 to 9, the EX trainer kits, Alternate Art Promos, Nintendo and Wizards Black Star Promos, First Partner Pack, Miscellaneous Promos, Best of Game, Yellow A Alternate, and McDonald's Collection 2021. The other 8 are one set, Mega Evolution Energy, which the matcher declines because the metadata spells it Energies and a singular-to-plural rule is not worth the wrong matches it would make elsewhere. Japanese sets carry release dates for only 2.99% of Japanese cards, so the set-age factor is reported on English cards only. The set-level restatement groups the matched panel by `cards.set_name`, keeps sets with 20 or more members, and correlates set age against the equal-weight mean of the members' forward returns; that grouping is the one figure in this module that moves with the duplicate-id tie-break above, so it is reported as a direction rather than a measurement. EXCLUSIONS, stated because an earlier draft of this block wrongly said there were none: day 20706 is one of the partial catalogue dumps, so requiring a price on it drops 2,441 of the 43,800 products priced on both endpoints, 5.57%, and those are cheap, with a median start price of $0.11 against $1.89 for Panel A. Their median 23-day return is 0.00%, the same as Panel A's, so the exclusion moves level figures rather than the central finding, but it is a price-correlated exclusion and not a neutral one.