Every number on a PokemonMarketWatch forecast page traces to a specific input on a specific date. This page documents how the model works, what each input contributes, where it comes from, and what the model can't do. Read this once and you will be able to audit any forecast we publish.
What the predictor is
PMW's market predictor is a transparent, versioned, heuristic forecasting engine. For each tracked Japanese set we publish three forecasts daily: a 30-day, 90-day, and 12-month directional call with an explicit confidence band. Every forecast is a forecast, not a fact, and is framed as such on every page.
The model is rule-based and human-readable, not a black-box neural network. We chose that deliberately. The Pokemon TCG market has a small number of repeating structural drivers (the four-phase reprint cycle, the JP-to-US lag, FX, supply-event shocks). A heuristic engine that names and weighs each driver lets readers audit and dispute our output, which is the whole point of publishing it.
The six input factors
Every forecast is the sum of six factor contributions plus a model-version-stamped intercept. Each factor's contribution to the final number is preserved in the database and shown on every forecast card so the math is never hidden.
1. Phase bias
Where the set sits in PMW's four-phase reprint cycle, which is documented in the Print Run Timeline:
- Phase 1 — Pre-Launch Lottery. Allocation constrained. Secondary at 200 to 400 percent of MSRP. Sentiment-led.
- Phase 2 — Launch Hype. First 30 to 90 days post-release. Secondary at 200 to 300 percent of MSRP. Reprint timing is the dominant variable for the next 6 months.
- Phase 3 — Reprint Trough. Reprints absorbed. Secondary at 80 to 150 percent of MSRP. Sets in this phase tend to drift sideways or down for 6 to 18 months.
- Phase 4 — Zetsuban Climb. Out of print. Prices compound upward on declining supply. The vintage tail of the Japanese market.
Each phase carries a base bias for each forecast horizon. Phase 4 carries the strongest positive 12-month bias. Phase 3 carries the only negative 30-day bias. Pre-launch Phase 1 carries the strongest positive 30-day bias because lottery hype concentrates into a short window.
2. Sentiment bias
Pulled from the sentiment_snapshots table's aggregate row for the set. Sentiment is normalized to a 0 to 100 hype index from four sources: Snkrdunk Magazine (0.30 weight), pokeca-chart listing velocity (0.30), Twitter JP (0.20), and trade-press supply events (0.20). Higher hype contributes positively at short horizons, but the weight tapers fast: at the 30-day horizon, hype gets a 0.20 multiplier; at 90 days, 0.10; at 365 days, only 0.03. Long-run prices revert toward fundamentals.
3. FX bias (JPY/USD)
Pulled from the fx_history table. We measure the trailing 30-day percent move of JPY versus USD. A weaker JPY makes JP-domestic sealed cheaper for international arbitrage buyers, which pushes JP-domestic prices up. PMW's coefficient for this is 0.4, the midpoint of the 0.3 to 0.5 elasticity range observed in Snkrdunk and Mandarake data. FX matters more at longer horizons because exchange-rate effects accumulate; we multiply the bias by 0.5 at 30 days, 1.0 at 90 days, and 1.5 at 365 days.
4. Supply-event bias
Pulled from the supply_events table. Each event has a base bias keyed off its type:
- Reprint announced: minus 6 points
- Reprint arrived: minus 8 points
- OOP confirmed: plus 8 points
- Restock: minus 3 points
- Lottery drop announced: plus 4 points
- JP price hike: plus 6 points
- EN price hike: plus 4 points
- MSRP change: plus 3 points
Set-specific events fade fast (60-day linear decay, full weight). Market-wide structural events (TPCi-Excell acquisition, Collectors-Beckett consolidation, Walmart purchase limits, JP My Number Card rollout) fade slowly (365-day linear decay, 0.35 weight) because they shift the base case for every set, not a single set.
5. JP-leads-US bias
Active only when forecasting the US side of a paired set. PMW's empirical research finds a 30 to 60 day pack-price lag JP-to-US for modern sets, with a 60 percent magnitude passthrough. The bias is strongest at the 90-day horizon (matches the empirical 60-day lag window) and tapers at 30 days (too early) and 365 days (too late).
6. Macro overlay
A small positive structural baseline reflecting the documented sustained undersupply in the global hobby channel: Millennium Print Group's capacity expansion does not complete until late 2028; TPCi's vertical integration with Excell Brands tightens hobby allocation; mass-retail limits cap scalper supply. The overlay is small (+0.5 at 30 days, +1.5 at 90 days, +4.0 at 365 days) and will be revisited when structural conditions change.
How confidence is computed
Confidence is a 0.00 to 1.00 number, multiplied across three sub-factors:
- Data density. How many sentiment snapshots have accumulated in the last 30 days. We need 30 snapshots for full credit; below that, confidence scales linearly. A pre-release set with no scrapers ingesting it gets a low score and an honestly-wide band.
- Release status. Post-release sets are scored 1.0, pre-release sets with a known date 0.6, sets with no announced release date 0.4. Forecasting before there is anything to anchor against carries inherent uncertainty.
- Horizon. 30 days carries 1.0, 90 days 0.85, 365 days 0.60. Long-horizon forecasts are mathematically less reliable.
The confidence number drives the band width on the forecast card directly: low confidence produces a wide band. A 5 percent point-estimate move on a 0.05 confidence forecast renders as a band from minus 25 percent to plus 35 percent. That is intellectually honest.
Where each input actually comes from
| Input | Source | Update cadence |
|---|---|---|
| Phase classification | Calculated from canonical_products.release_date_jp + recent supply_events | Daily |
| Hype index | sentiment_snapshots aggregate row, sourced from Snkrdunk Magazine RSS, pokeca-chart price API, TwitterAPI.io watchlist, and the supply-events stream | Daily 18:00 local |
| JPY/USD rate | fx_history populated by exchangerate-api.com | Daily 18:00 local |
| Supply events | supply_events populated by pokemon-card.com/info scraper, Snkrdunk Magazine RSS, Twitter JP scraper, manual editorial input for major TPCi/distributor events | Continuous (scraper) + as-disclosed (editorial) |
| JP price moves | price_snapshots from Yuyu-tei, Snkrdunk, Mandarake, Surugaya, pokeca-chart | Daily |
Model versioning and backtesting
Every forecast row in predictions_snapshots carries a model_version tag (currently pmw-v1.0-heuristic). When we ship updated weights, old forecasts retain their original version stamp. This lets us run head-to-head backtests across model versions and publish the comparison, which is a commitment no competitor in the Pokemon market currently makes.
We do not yet publish backtests because daily forecasts started accumulating on launch day. We will publish the first formal backtest 90 days after launch, with rolling updates quarterly. If a model version underperforms its successor, we will say so on this page and roll it back.
What this predictor does not do
- Individual card forecasts. The first version operates at the set level. Per-card forecasts require per-card price-history density that PMW does not yet have. They are on the roadmap, not in production.
- Real-money trading signals. Forecast pages are educational and analytical. PMW is not a registered broker, dealer, or investment advisor. Forecasts are not investment recommendations.
- Account for individual seller dynamics. The model forecasts aggregate market direction. It cannot predict whether a specific seller will price above or below the band.
- Predict unprecedented events. A new TPCi distribution policy, an unexpected fire at Millennium Print Group, a Japan-specific regulatory change, an unannounced lottery — none of these are in the input data until they happen. The model adapts on the next snapshot, but it does not foresee.
Auditing a specific forecast
Every forecast card shows the contribution of each of the six factors to the headline percent-move estimate. If a forecast looks wrong, you can decompose it into the factors and trace each one back to the underlying snapshot. The inputs_payload column on every predictions_snapshots row preserves the raw inputs we consumed, so editorial pieces (and the eventual backtest reports) can replay the model against any historical day.
If a forecast turns out badly wrong in hindsight, that is not a thing we hide. The methodology page version, the model version, and the inputs payload give us a full audit trail. Where competitors lean on opaque "AI-powered composite scores from six market indicators" framings, PMW publishes the indicators, the weights, and the formulas.
Methodology version history
- pmw-v1.0-heuristic (live, May 23 2026 onward) — initial six-factor model. Phase bias coefficients derived from the four-phase reprint cycle. Sentiment weights derived from JP secondary-venue cleanliness analysis. FX elasticity midpoint from observed JPY-weakness vs. JP-sealed-price correlations 2021 to 2026. JP-leads-US passthrough from documented pack-price-hike lag (October 2022 to December 2022 canonical precedent).