Vintage Backcast Series
How we estimate vintage card values for the years before clean measured data exists (1999 – 2022). Why the values are Backcast always shown separately from measured data, and never blended into a measured index trace.
The cleanest free historical price source for trading cards is TCGplayer's archive (via TCGCSV), which reaches back to roughly 2022. eBay sold-listing data has only ~90 days of public retention. PriceCharting's API offers no historical pull. Auction-house results pre-2010 are sparse and modeled even by their publishers.
Rather than pretend the gap doesn't exist, we explicitly model the missing years and label the result. A user who wants the 1999 → 2026 picture of Base Set Charizard PSA 10 should be able to see one — just not be misled into thinking the pre-2022 portion is the same kind of number as the post-2022 portion.
The GPT-5 backcast call receives, per (card, grade, year):
- Landmark sales — every public auction-house result we have for this card-grade pair, plus any widely reported private sale (Logan Paul Charizard 2022, etc.).
- PSA pop snapshots — historical population counts where we have them, otherwise the present count and a reverse-engineered growth assumption documented per row.
- Narrative context — pandemic spike, set release, anniversary print runs, major influencer events that moved the market.
- CPI normalization — US CPI series so the model can sanity-check that the implied real return is plausible.
- Comparable cards — for a target card, the backcast for two or three other cards in the same era that already passed quality review.
Every backcast row stores:
- llm_model_version — exact GPT-5 snapshot used
- prompt_hash — SHA-256 of the prompt template
- methodology_version— this page's version
- inputs — the full list of inputs the model received, stored alongside in
csm_vintage_backcast_inputs
Re-running the same inputs through the same model snapshot must reproduce the same numbers. Tests assert this property for every published row.
For every (card, grade) we have at least one anchored measured point — typically the post-2022 weighted-median fair value. The backcast is calibrated so that its 2022 endpoint matches the measured 2022 starting value within a documented tolerance band. Calibration drift exceeding 8% triggers a quality review.
Every place a chart shows both backcast and measured data the two traces are visually distinct: the backcast trace is dashed and uses a different color (cyan-violet vs the measured cyan). The legend always identifies it as "Backcast — modeled".
In tables and stat blocks, backcast values appear in their own dedicated section under the Backcast badge — never mixed into a row that contains measured values for an adjacent year.
- Not a fair-value estimate for buying or selling today. Use the measured per-grade page for that.
- Not an investment recommendation. Backcast is research only.
- Not blended into any index, market cap, or macro figure.