Fair Value Methodology

Every graded card on TCGBerg carries a fair value: our best estimate of what that exact card, in that exact grade, is worth today. It is built from actual completed sales of professionally graded copies, not asking prices, not listings, and not a dealer markup. The estimate is recomputed daily, expressed in US dollars, and published with a sales-evidence rating that tells you how much market evidence stands behind it. This page explains how the number is built, what we do to keep bad data out of it, and what it does and does not promise.

Maintained by the TCGBerg research team · Last updated August 20, 2026

The Atom Model

The unit of analysis is the atom: the (printing, grader, grade) tuple. A printing is a specific physical print run of a card, so "Charizard Base Set 1st Edition Shadowless" is a different printing from "Charizard Base Set Unlimited" even though they share a card identity. A grade is the quality level a grading company assigned.

Why this granularity? Because the market values atoms, not cards. A 1st Edition Shadowless PSA 10 and an Unlimited PSA 10 share a card name but are economically distinct: the supply, the demand, and the price discovery have little overlap. Aggregating them into one "Charizard PSA 10" average would obscure both the individual prices and the scarcity premium that separates them.

A single card can therefore correspond to dozens of atoms. Fair value runs once per atom per day, and every page on TCGBerg either renders an atom's metrics or aggregates atoms.

Today we publish fair values for PSA grades 6 through 10. Each grade is valued entirely on its own sales. We track BGS and CGC populations, and they appear in our population data, but we do not yet publish a BGS or CGC fair value; extending the engine to other graders means treating each company's scale on its own terms rather than assuming a PSA 9 and a BGS 9 are the same asset.

Built From Completed Sales

Fair value is grounded in transactions that actually happened. We continuously collect the prices real, professionally graded copies have sold for, and the estimate is derived from that sales record. Asking prices never enter it. A listing is what somebody hopes to get; a sale is what somebody got.

The window is measured in sales, not days. The engine asks "what were this card's most recent transactions?" rather than "what happened in the last N days", because that is how dealers and collectors actually think about a thinly traded asset, and because a fixed calendar window returns nothing at all for a card that trades a few times a year.

Sales in other currencies are converted to US dollars at the official daily reference rate for the date the value is calculated, so a European auction and a domestic sale enter the same sample on the same basis.

Every published value records which sales produced it, so a figure on the site can be traced back to the transactions behind it rather than taken on faith.

Recent Sales Count For More

Card markets move, so a sale from last month tells us more about today than a sale from two years ago. Recent sales are weighted more heavily, and the weighting is by calendar age: what matters is how long ago a sale happened, not its position in a queue. Two sales on the same day count the same, and weight falls away smoothly with age rather than at a cliff edge, so no single sale ages out of relevance overnight.

How quickly that weighting decays adapts to the card. A card trading several times a month gets a tight, responsive window, because there is enough recent evidence to trust it. A card trading a few times a year gets a longer memory, because insisting on recency there would leave nothing to work with.

The point estimate is not a single formula. Several estimators run over the screened sample in parallel, each capturing something different: short-run momentum, a robust central value that resists skew, an anchor to the most recent trading window when one is deep enough, and a trend projection that only contributes when a trend is statistically real rather than noise. Their contributions are then blended, and the blend adapts per card to how scattered its prices are, how densely it has been trading, how strong any trend is, and how long it has been since it last sold. A noisy sale stream leans on the robust estimator; a deep recent sample leans on recent trades; a quiet card leans conservative. The result is designed to be stable through one-off flukes yet responsive to a genuine change in the market.

Not Every Recorded Sale Is A Clean Signal

Some individual sales are misleading: a damaged copy, a mislabeled listing, a lot of several cards sold as one, a price typo, or a transaction that was never an arm's-length sale. Screening these out is most of the work, and it happens in layers before any estimate is computed.

Cross-grade coherence. A card's grades form a ladder, and a recorded top-grade sale priced like a heavily played copy is almost always a mislabel rather than a bargain. Sales that sit incoherently against the neighbouring grades of the same card are removed.

An asymmetric fence. We then test each sale against the going rate for that atom using a robust measure of the spread, one that is not itself dragged around by the outlier it is trying to catch. Crucially, we do this asymmetrically, because the two sides are not equal:
- A sale far below the going rate is almost always a data problem, so we are strict about setting those aside.
- A sale far above the going rate can be perfectly genuine: a bidding war, an exceptional copy, a real repricing. So we do not discard it reflexively. It is kept at full weight when something corroborates it, either other sales in the same window agreeing with it, or comparable cards moving together: when a card's peers are all climbing, a high sale is evidence of a real move rather than an anomaly, whereas the same sale against flat peers is a lone spike.

Reduced influence rather than a verdict. An uncorroborated high sale is not simply deleted. It is admitted at reduced influence, so it still informs the estimate in proportion to how far out it sits and how much support it has. This matters because rejection is a binary answer to a question that is rarely binary: a genuine repricing usually shows up first as exactly one sale nobody else has matched yet, and throwing it away means the estimate is always the last to know. Letting it in quietly lets the value begin to move without letting one trade set it.

A percentile trim. What survives is finally clipped at the extreme tails rather than deleted, which neutralizes a residual distortion while preserving the sample size. That matters most on thinly traded cards, where every observation counts and throwing one away costs more than clipping it. Whether the trim had to intervene is recorded and feeds the sales-evidence rating.

The goal throughout is to reflect what the card genuinely trades for, not to chase noise, and not to quietly delete the evidence that the market moved.

Rarely Traded Cards Borrow Signal From Their Neighbours

Some cards sell often; most sell rarely. When a card trades too infrequently to read cleanly on its own, we borrow signal from closely related cards: comparable cards in the same set, at a similar grade and price level, whose collective movement tells us which way that corner of the market is going.

This is used in two places.

Older comparable sales are re-expressed in today's terms. If a card's most recent sales are months apart, those older prices are stale evidence about today. Each one is adjusted toward current conditions by how far comparable cards have moved since that sale happened, in a measured, dampened way with a hard limit on how far any single sale can be moved. Every stage downstream then judges prices that are consistent with the same date.

Cards that have gone quiet are re-anchored. If a card has not sold in a long time, its last price can drift badly out of date while the wider market moves. Rather than leaving a frozen number on the page, the estimate is nudged toward where comparable cards are trading now. The nudge is dampened and capped, never a wholesale repricing, and a value that has been materially adjusted this way carries a weaker sales-evidence rating to say so.

Both mechanisms are deliberately conservative. They exist so a scarce card shows a sensible current estimate instead of a price from the last time it happened to sell, not to manufacture movement a card has not earned.

Keeping The Grade Ladder Coherent

A higher grade of a card should not be worth less than a lower grade of the same card. When the underlying sales would produce that inversion, something is usually wrong with one side of it rather than with the market.

The ladder is reconciled as a whole, not pair by pair. Walking up a card's grades fixing each crossing against the one below it sounds sufficient and is not: a large share of real crossings are not between adjacent grades, and repairing one pair routinely creates a new violation further up. So the whole PSA 6 to 10 ladder for a card is fitted at once into an ordered sequence, in proportional terms because prices are multiplicative, and weighted by how much evidence each grade carries so a well-evidenced grade holds its level and a thin one moves toward it rather than the other way round. Two grades are also never left at exactly the same price, since a ladder that reports a PSA 9 and a PSA 10 as identically valued is telling you nothing about either.

Where one grade is thin and its neighbour well evidenced, the thin side is corrected. Where both sides are genuinely well evidenced, a crossing is treated with more care: occasionally the market really does pay more for a lower grade, and flattening that would be inventing data rather than correcting it.

A related guard runs in the other direction. A thin grade priced implausibly far above the grade below it, on very few sales, is not treated as a discovery; its sales-evidence rating is demoted so nothing downstream leans on it.

Scarcity is allowed to lift a grade, but only on evidence. Where the grade above is genuinely much scarcer than the one below and there are enough slabs at that grade for the comparison to mean anything, the ladder reflects that scarcity premium. Both conditions are required and both are read from published population; a grade that fails either is left exactly as the sales left it.

Every Value Comes With A Sales-Evidence Rating

No estimate is certain, so each fair value is published with a sales-evidence rating that says how much market evidence stands behind it. The rating is read from an internal 0-100 score that combines five things:
  - How many sales back the estimate.
  - How recent the most recent one is.
  - How regularly the card trades, rather than clustering and then going silent.
  - How closely the sales agree with each other.
  - Whether the screening steps had to intervene on the sample at all.

The score maps onto the five sales-evidence levels shown across the site: 80 and above is Very strong, 60 to 79 Strong, 40 to 59 Moderate, 20 to 39 Limited, and below 20 Sparse. Those levels are what the Sales evidence marks on every card page show.

The rating is load-bearing rather than decorative. Indices down-weight thinly evidenced atoms, the deal screener refuses to flag a discount measured against a value it does not trust, and a value resting on too few or too old sales is capped at the bottom of the scale: it still publishes, but it never advertises authority it has not got.

One honest caveat: thinly traded cards carry more uncertainty. Even a careful estimate on a card that sells twice a year has a wider margin than one on a card that sells weekly. The sales-evidence rating exists so you always know which kind of number you are looking at.

Raw (Ungraded) Values

Raw cards are shown per condition, and the number you see is one of two things.

Where there are enough completed ungraded sales for a card in a given condition, we model a raw fair value using the same engine and the same sales-evidence rating as the graded side. Where there are not, we show a market price quoted by an ungraded price aggregator instead. The card page distinguishes these visually, so you can always tell a modelled value from a quoted one.

Raw coverage is the mirror image of graded coverage, and that is expected rather than a gap. Bulk and mid-tier cards trade raw constantly, so they model well. Marquee chase cards in near-mint condition rarely sell raw at all, because a copy that good gets graded instead. So the cards with the deepest graded data are usually the ones with the thinnest raw data, and a quoted rather than modelled raw price on a headline card is the normal case, not an error.

Raw atoms have no graded population, so they carry a value but no market cap, and they do not enter index or market cap rollups.

What Fair Value Is Not, And Where It Is Weakest

Fair value is an estimate, and it is worth being precise about the limits:
  - It is not an appraisal or a certified valuation.
  - It is not an offer to buy or sell at that price, and not a guarantee that any copy will sell for it. What a specific card fetches depends on its eye appeal, the marketplace, the timing, and the buyer.
  - It is not investment advice. It is a market reference to help you understand where a card is trading; the decisions are your own.

Known weaknesses we would rather state than hide:
  - Thin samples. Even a single recent sale produces a value, with an appropriately weak sales-evidence rating.
  - Some grades are valued without any sales of their own. On a card where neighbouring grades have traded but one rung has not, that rung can be given a value derived from the rest of its own ladder and from how scarce it is, rather than being left blank. These are the least certain numbers on the site and they say so: they are published at the bottom of the sales-evidence scale. We would rather show a clearly-labelled weak estimate than an empty cell that implies a card has no value, but treat them as an indication of where a rung sits, not as a price. Where a card has no usable sales anywhere on its ladder, nothing is published and it shows an empty state.
  - A sale-count window can reach a long way back. On a card that trades rarely, the most recent transactions may span years. The peer adjustments above exist to counter that, but a very quiet card is still the hardest case on the site.
  - Sale venues are not weighted. An auction-house hammer price and a marketplace sale enter the sample on equal terms; venue premiums are not deflated.
  - Ambiguous attributions are excluded. A sale we cannot confidently tie to a specific printing is left out rather than risk polluting the wrong card.
  - Graded coverage is PSA-first. Other graders' populations are tracked, but their sales do not yet produce their own fair values.

Frequently asked

Is fair value the same as the last sale?

No. It is an estimate built from a screened window of recent transactions, not a single price. One sale well above the prior fair value moves the output modestly, exactly because one trade should not dominate an asset's value. If a single sale did move it sharply, that would make the number less useful, not more current.

How often does fair value update?

Daily. After the overnight ingestion completes, every atom's value and sales-evidence rating are recomputed from the latest sales data. Recomputing the same date twice produces the same answer.

Why did the fair value not move after a sale I saw?

Three possibilities. The sale may still be working through ingestion. It may have been screened out as incoherent against the card's other grades, or as an unusually low price with nothing corroborating it. Or it may simply be one observation in a window of many, in which case it moved the estimate a little rather than a lot.The sales behind a value are recorded, so a value is always traceable.

Does TCGBerg use machine learning?

Not in the sense of a model that prices a card end to end, and we would rather explain the distinction than give a one-word answer. A card's value is not predicted by a trained model; it is computed from that card's own screened sales by transparent statistical steps.But some of the corrections around that core do use coefficients fitted offline on historical sales, most notably the one that estimates a grade with no sales of its own from the rest of its ladder. Those are fitted on held-out data, validated against sales the fit never saw, then published as fixed constants and applied as stated numbers. Nothing is re-fitted at request time, no model output goes out unaudited, and every published value keeps a record of how it was reached. The commitment is not the absence of fitting, it is that a five-figure valuation always has a derivation you can follow rather than a black-box number.

How is sales evidence different from accuracy?

Sales evidence is a measure of the evidence itself, computed from the sample's size, recency, regularity, agreement, and cleanliness. It says how much support the estimate has, not how close it will land to the next sale. Accuracy is tracked separately against later transactions, and the two generally move together.

Why clip extreme prices instead of deleting them?

Deleting discards information, and it hurts most exactly where samples are thinnest. Clipping an extreme price back toward the sample neutralizes the distortion while keeping the observation, and the sales-evidence rating still records that the sample needed cleaning.An unusually high sale that nothing corroborates is not deleted either: it is admitted at reduced influence, because a real repricing normally arrives as one sale before it arrives as several. Sales that are not merely extreme but incoherent, such as a top-grade sale priced like a damaged copy, are removed outright.

Why does a grade with no recorded sales still show a value?

Because a blank cell reads as "this card is worthless" when what it means is "nobody has sold this exact grade recently". Where the rest of a card's grade ladder has traded, that rung can be estimated from its own ladder and from how scarce it is.Those estimates are published at the bottom of the sales-evidence scale precisely so they cannot be mistaken for a traded price. Where a card has no usable sales at any grade, nothing is published.

Are international sales included?

Yes. Sales in other currencies are converted to US dollars at the official daily reference rate for the date the value is calculated, and historical values use the rate that applied at the time.

Do you publish fair values for BGS and CGC?

Not yet. We track BGS and CGC populations, and they appear in our population data, but fair value is currently published for PSA grades 6 through 10 only. Each grading company's scale means something different, so extending the engine means valuing each on its own sales rather than converting between scales.