Updated August 2026 · model version forecast-deep-v4.4 · retrained on every data refresh
CardStock treats trading cards like a market you can actually study. Every card gets a price forecast for 1 month, 6 months, and 1 year ahead, for the raw card and each graded tier, next to real graded price history. Weekly market reports round up the biggest movers and a public scorecard of how the model's own predictions have performed. None of it is financial advice.
What each card is comes from TCGplayer's catalog. What each card costs comes from two places: raw prices are TCGplayer's sales-backed market price, collected on every refresh; graded prices come from PriceCharting's per-grade sales history. Prices also police each other: a market price is only trusted while real sales stand behind it — a frozen price that disagrees with fresh sales elsewhere gets flagged, reviewed, and switched to the sales-backed source. When neither source has defensible data, the card shows no price at all.
Price momentum and volatility, the card's set and its price relative to set-mates, game-wide and hobby-wide trends, the card's printed characteristics and artwork, and the model's own trailing forecast errors per card and per set. Under the hood: gradient-boosted decision trees per game, tier, and horizon, trained on historical outcomes with held-out validation.
Every published forecast is saved, and once its date arrives it is graded against what the price actually did — misses stay on display on each card's chart. The weekly report publishes accuracy, bias, and how often prices landed inside the model's ranges, whether flattering or not.
New cards carry no forecast until they have at least two months of clean price history — we tested launch-window prediction and the honest answer was: not well enough to publish.