Never Buy the Small Dip: Cards Just Under Their Own High Made 4.52% Over Three Months, Cards At Their High Made 9.27%

The most common thing a collector does is wait for a small pullback. The card is $40, it was $44 last month, so you wait for $40 to become $38 and then you buy. We measured what that patience is worth. Say the population before any of it, because the population decides two of the numbers: 29,862 cards carry a recorded weekly price series, 5,431 of them have a reading on 27 January 2025, 4,993 of those were still being quoted in the last four weeks, and 2,679 join to exactly one English catalogue row. The 1,376 of those that produced at least one reading at $5 or more with 13 further weeks of readings ahead of it are this study, 79,786 card-and-week observations in all. Every one of them is scored against the mean forward return of every other card read in the same week in the same price band, so a rising market cannot pose as a finding. On that measure the small pullback is the worst of the seven states we measured, and the deep falls scored best.

Distance below the high is measured against the card's own running high in our recorded weekly prices, so it is a number a reader can read off the price history chart on any card page without a model. Forward return is the move over the next 13 weeks. The excess column is that return minus the mean return of every observation in the same week and the same price band, in points, which is the control that stops a rising market flattering every row.

Read the excess column top to bottom. It falls from the first row to the second, bottoms at 5 to 10 percent below the high, and then climbs as the discount deepens, rising through every band to 40 to 50 percent below before easing back a little in the last row. The second row is the lowest of the seven, and it is the lowest of the seven in every one of the six specifications further down this page. Two rows deserve their caveat printed next to them rather than in a footnote: 40 to 50 percent below rests on 836 observations across 95 cards, and 50 percent or more below rests on 369 observations across 34 cards. Those are small populations, and the last of them is 34 cards. Unlike the wider bands, neither of them changes sign across the six specs, but neither is a number to size a position on either.

One honest wrinkle about the excess column, because it decides how the first row should be read. Two thirds of all observations, 67.11%, sit within 5 percent of the high and 41.18% sit exactly at it, so that band is most of its own cohort and its excess is pulled toward zero by construction. Its +0.7 is therefore not the interesting number. The gap between the two top rows is, and the gap is stable whichever centre you use: 4.5 points on the means, 4.6 points on the medians. The excess distribution is skewed right, which is why the median excess is negative in five of the seven bands while the mean is not, and it is why the mean is the control figure quoted here and the median is quoted for the raw return.

9,278 small dip observations are not 9,278 independent facts. They are roughly nine overlapping 13 week windows on each of 998 cards, so an interval computed per observation would be far too narrow. Resampling cards rather than rows, 1,000 draws, the small dip's shortfall against the at-high state is 4.5 points with a 95% interval from 3.6 to 5.4 points. Card by card the picture is broad rather than concentrated: of the 753 cards with four or more weeks spent 5 to 10 percent under their high, 64.94% had a negative mean excess, and the median card in that group came in 5.8 points behind its cohort. This is not a handful of cards carrying a result.

The natural dismissal is that a small dip is simply the first week of a collapse, so the penalty is really the collapse arriving later. We tested it two ways and it does not rescue the dip. Delete every observation that is followed anywhere later in the window by the card falling 30 percent or more, and the small dip still trails: excess -3.3 on 8,431 observations against +0.9 for the at-high band, with 17.02% clearing the round trip against 22.70%. Delete every card that ever sat 30 percent or more below its high at any point, which throws away 195 cards and leaves 1,181, and the small dip still trails: excess -3.5 on 8,365 observations across 829 cards against +0.8 at the high, 16.69% clearing against 22.71%. The penalty survives both deletions almost intact.

The lagged entry specification is the one worth a sentence. This repo has been burned before by a signal that shared a price with the denominator of its own forward return, which manufactured a bounce out of a single bad print. Here the artifact would work in the dip's favour rather than against it, because a low print would create both the apparent dip and the apparent recovery. Reading the state in one week and buying in the next breaks that link, and the penalty is 3.4 points instead of 3.8. The 26 week column says the penalty grows with the horizon rather than washing out, and it is one number on one horizon rather than a trend.

The mechanism is visible in what each state looks like backwards rather than forwards. A card within 5 percent of its high set that high a week ago at the median and has risen 15.19% over the previous 13 weeks. A card 5 to 10 percent under its high set that high eleven weeks ago and has gone nowhere since, a median trailing return of +0.94%. The small pullback is not a rising card briefly on sale. It is a card whose rise ended about two and a half months ago and has not yet fallen far enough for anyone to call it cheap.

That is consistent with what Momentum Does Not Persist in Pokemon Cards found on a much larger population, 52,041 cards, when it retracted an earlier momentum result. This study quotes that piece rather than re-testing it: nothing here is a momentum backtest, the pool is 1,376 survivors against its 52,041, and a survivor pool is exactly the error that piece was written about.

If the bottom of the table is the good state, the next question is when to step into it. Take every first crossing of 30 percent below the running high where the crossing price was $5 or more: 196 episodes on 196 cards. Only 15.31% of them bottomed in the week they crossed, which is another way of saying 84.69% had further to fall. The median episode reached its low 6 weeks after the crossing week, 64.80% of them four weeks or more later, and the median further fall from the crossing price to the low was 10.08%, taking the median trough to 38.71% below the prior high. Restricting to the 178 crossings with at least 13 more weeks of readings after them, which removes the episodes that ran out of readings rather than out of downside, gives 12.92% bottoming in the crossing week, a median 7 weeks to the low, and a median further fall of 10.70%.

Now the part that decides the instruction. A card that has been down for a while is also usually down further, so an apparent reward for patience could just be a reward for depth wearing a calendar. Hold depth fixed and the calendar still separates. Inside the single band of cards sitting 30 to 40 percent below their high, the first four weeks of the fall came in 1.5 points below their cohort and the fifth week onward came in 5.4 points above it, a difference of 6.8 points with a 95% interval from 3.0 to 9.6 points, resampling cards. The clearing share tells the same story in the units a reader cares about: 19.58% against 32.77% at the 24.14% cost floor, and 29.93% against 46.73% at the 14.94% floor.

Pooled across every depth at 30 percent or worse the same split reads +0.8 against +6.6, on 926 and 2,213 observations, and the pooled difference is 5.8 points with an interval that just touches zero because 124 of the weeks-one-to-four observations sit at 40 percent or deeper on only 52 cards and pull the thin end around. The depth-controlled version above is the one to carry, and it is the one that survives. For scale, an observation not in a 30 percent drawdown at all cleared the 24.14% floor 21.99% of the time, against 35.61% for the aged fall.

All three readings come off two numbers a reader already has: the card's own highest recorded weekly price and its price now. Waiting for a 5 to 10 percent pullback is the one state to stop paying for; on this evidence you either buy at the high or you wait for a real fall. A card that has just crossed 30 percent below its high is not yet a purchase, because five in six of them had not finished falling, the median episode fell another tenth before reaching its low, and that low arrived six weeks after the crossing week at the median. A card that has held 30 percent or more below its high for five weeks or longer and is still there is the state that scored best, and it is a real fall rather than a rounding error: the median such observation sat 39.74% under its high and had been under for 13 weeks.

As of the last complete week in our recorded prices, 21 September 2026, 1,360 of the 2,679 continuously quoted cards carried a price of $5 or more, which is the population for the four counts that follow. 845 of them sat within 5 percent of their own high, 186 sat 5 to 10 percent below it, and 68 sat 30 percent or more below, of which 43 had been there for five straight weeks or longer. The table below is the 11 of those 43 whose high is not a single print: their highest weekly price was matched within 2 percent in three or more separate weeks, so the drawdown is measured from a level the card actually held. This is a list of cards in a state, not a list of recommendations, and the state's edge is an average across a rising window.

Three things about that table a reader should hold against it. Ten of the eleven come from the Sun and Moon era and nine of the eleven are secret rares, which is what a pool of continuously quoted 2017 to 2019 cards looks like rather than a finding about secret rares. Nest Ball set its high on the very first week we hold, 27 January 2025, so its real peak may sit before our window opens entirely; every other row set its high inside the window, and that row is also the only one whose set has no page on this site yet, which is why it is the one set name in the table without a link. And the other 32 of the 43 are left out here only because their high was matched in fewer than three separate weeks, which makes the measured depth harder to trust rather than the card worse.

Two cost floors appear in every table above and they are never averaged. 14.94% is the gross move needed to break even on fees and postage alone, the figure How Long Do You Have to Hold a Pokemon Card modelled for a seller. 24.14% adds an 8% gap between the quote you buy at and the mark you are measured against, which is the floor our own pick engine gates on and the one Momentum Does Not Persist in Pokemon Cards used. A buyer who also has to sell faces the second one. The corpus carries both because they answer different questions, and every share in this study is printed at both so no reader has to guess which was used.

Source: the recorded per-card weekly price series this site accretes, which is the TCGplayer market price read weekly on a Monday from 27 January 2025, thickening to daily readings from late July 2026. Every reading is snapped to its Monday and the last reading in a week is that week's price, giving 87 weekly slots from 2025-01-27 to 2026-09-21; the partial week beginning 2026-09-28 is dropped. Two of the 87 slots are thin in the source (the weeks of 2026-08-10 and 2026-08-17 carry 7,925 and 5,313 of the 29,862 series), so continuity is required at the ends rather than in every slot: a series enters the pool if it has a reading on 2025-01-27 and a reading in the last four weekly slots. Funnel: 29,862 series, 5,431 with a 2025-01-27 reading, 4,993 still quoted in the last four slots, 2,679 joining exactly one English catalogue row. Catalogue join: 50,504 TCGplayer product ids appear in the catalogue and 1,215 of them are claimed by two rows each, almost always the same set spelled twice, so rows are filtered to English first and the 243 product ids still doubled after that are resolved to the lower card id, which is the rule the index builder uses. An observation is a card and a week where the card is priced at $5.00 or more and a reading exists 13 weekly slots later, which gives entry weeks from 2025-01-27 to 2026-06-22, 74 slots, and 79,786 observations on 1,376 cards. Price bands are $5 to $10 (26,899 observations), $10 to $25 (24,280), $25 to $100 (21,223) and $100 and up (7,384). Distance below the high is the week's price against the highest weekly price recorded for that card up to and including that week, so the first reading of every series is its own high by construction; the fourth specification in the table above re-runs everything requiring 13 weeks of prior history and the result moves by half a point. Excess is the observation's 13 week return minus the mean 13 week return of every observation in the same weekly slot and the same price band, reported in points to one decimal. Every interval quoted is a 1,000-draw bootstrap that resamples cards rather than observations, because each card contributes about nine overlapping windows and a per-observation interval would be several parts too narrow. Cost floors: 14.94% is 1 divided by 0.87 minus 1, fees and postage on the sale only; 24.14% is 1.08 divided by 0.87 minus 1, adding an 8% entry gap, and it is the floor the pick engine gates on. Both are printed and never averaged. A weekly market price is neither an asking price a buyer can hit today nor a graded comp; it is a mark, and a 13 week return computed across marks is not a realised return. No set released inside the window can appear, because every card in the pool was already quoted on 2025-01-27, which is what rules out a launch collapse being read as a drawdown. Three published studies on this same series carry three pool sizes, and the reasons are mechanical rather than a disagreement: The Drawdown Book used 2,417 cards with no catalogue join and no forward-window requirement, How Far Do Pokemon Cards Fall and How Long Do You Have to Hold a Pokemon Card used 2,766 on a longer daily span, and this study's 1,376 is what remains after one row per product id, a $5 floor and 13 forward weeks on every observation. The Drawdown Book's unconditional +9.7% median and 76% positive over 90 days are near this pool's +8.35% and 73.98% over 13 weeks, which is agreement in the neighbourhood rather than a replication, because the pools and the horizons are not the same.