How we price a card

A working tour of the model behind every HyperPotion call — what goes in, what comes out, and what we don't pretend to know.

The question we're answering

HyperPotion isn't trying to predict where a card's price will be next week. We're answering a narrower, sturdier question: what is this card actually worth right now, given everything we can observe about supply and demand — and is the market trading above or below that number?

Think of it like a bond yield. There's a model number, and there's what someone's actually paying. We measure the gap. The gap is the call.

What goes into fair value

Four real drivers do most of the work. Pull cost is how much sealed product you'd have to rip to find one copy — the rarer the chase, the steeper the floor. Character demand is the reason Pikachu costs more than Pansage at the same rarity slot, year in and year out. Grading economics tells us that a PSA 10 is worth more on a card where 10s are hard to make, and less where the gem rate is generous. And trade pace — how often a card actually changes hands — sets very different price floors for a card that trades ten times a week versus one that moves once a month.

Everything else — set name, hype cycles, who tweeted what — is noise we let the data filter.

The two-tier model

Cards land in one of two buckets. Tier A covers cards with enough PSA 10 sales history that we can anchor fair value directly to graded comps — the slab market is doing the heavy lifting for us, and we just follow. Tier B covers everything else: the cards where graded volume is thin, so instead of pretending we have data we don't, we price them off the underlying fundamentals — pull cost, character premium, gem rate, velocity.

The split exists because being honest about what we know is more valuable than pretending one model fits everything. On the cards it trained on, the model explains 92% of price variation — and, more to the point, 85% out-of-sample, where we hold back a whole set, predict it cold, and check the damage (leave-one-set-out). That second number is the bar we hold it to, because it's the one that says anything about cards the model has never seen.

From fair value to a verdict

The residual is the gap — market price minus fair value, as a percent. A 10% gap on a Hyper Rare doesn't mean the same thing as a 40% gap on a Special Illustration Rare; different rarities have different normal volatility, so a flat threshold would lie to you. Per-rarity cutoffs mean a Strong call always reads the same way: this is a big number for this kind of card. Lean and Watch are softer signals — directional, worth a look, but not high confidence.

Why we surface liquidity

A 30% discount you can't exit is not a discount. Every card gets a tag — Hot, Active, Watching, Thin — based on how often it actually trades. Investors can filter to the cards they can move in and out of; collectors who just want a deal can ignore it entirely. Same data, two reads.

What we don't claim

We don't predict next week's price. We don't price one-of-one promos or hand-cut error cards. We don't know which character is about to get a TV reboot. What we do: measure what's there, in the data, in the trades, every day, against a model tested out-of-sample on 1,128 cards. The rest is yours to decide.