How Our Price Predictions Work

A plain-English look at the data, the statistical signals, and the machine learning model behind every One Piece TCG price projection on WealthFamePower.

The short version

Every English-language One Piece TCG card with a TCGPlayer listing gets a daily price projection over two horizons, 7 days and 30 days. Each projection has a direction (up, down, or flat), an expected percentage move, and a confidence score. You can see them all on the price predictions page. These are statistical projections, not financial advice.

Where the price data comes from

Every prediction starts with TCGPlayer market price data, sourced through the free TCGCSV API. TCGCSV maintains daily snapshots of TCGPlayer's full catalog including market prices, low, mid, and high price tiers, and foil pricing.

WealthFamePower runs a sync job once per day at 4:00 PM Eastern. That job fetches the latest snapshot, updates each card's current price, appends a row to the price history table, generates a fresh round of baseline predictions, and checks user price alerts. Because the sync runs daily, price history accumulates by one day every day.

The full daily history starts in late January 2026 and has been collected continuously since then, giving the models a contiguous window of roughly 111 days to learn from. Every additional day of clean data makes the models incrementally more reliable.

The baseline statistical model

The baseline model is fully transparent and rule-based. For every card variant it computes three signals from the recent price series, combines them, and turns the result into a direction and an expected move.

1. Momentum

How much has the price moved over the lookback window? Cards trending up keep trending up more often than not over short horizons, so a strong recent gain pushes the prediction up. The signal is capped so a single outlier day cannot dominate.

2. Mean reversion

How far is today's price from its recent average? When a price spikes far above the average it tends to drift back down, and vice versa. We compute a z-score of the current price relative to the window mean, then flip the sign: a price well above the mean creates a "down" signal and a price well below creates an "up" signal.

3. Volatility

How choppy has the series been? Volatility does not push the direction one way or the other. Instead it shapes the expected magnitude (a volatile card can move further) and it discounts confidence (a volatile card is harder to forecast).

We average momentum and mean-reversion into a single ensemble score. If the score is above a small dead-band threshold the direction is "up"; below the negative threshold it is "down"; anything in between is "flat." The expected magnitude is then sized using the recent daily volatility and scaled by the square root of the horizon, which is the same approach used for sizing expected moves in equity options.

The machine learning model (beta)

The ML model is a LightGBM gradient-boosted ensemble. It is actually two models stitched together for each horizon: a classifier that picks direction (up, down, or flat) and a regressor that estimates the percentage magnitude of the move.

For each card it sees about 30 features per prediction:

  • Price features: lagged prices at 1, 3, 7, and 14 days, rolling mean and standard deviation over 7 and 30 days, and returns over 1, 3, and 7 days.
  • Card metadata: rarity, kind (Leader, Character, Event, Stage), set, number of colors, cost, power, life, counter, and the age of the set.
  • Social signals: recent Reddit mentions, a derived social-attention score, and a spike percentage that catches sudden surges in chatter.
  • Baseline signals: the momentum, mean-reversion, and volatility readings from the baseline model are themselves features, so the ML model can decide when to lean on them and when to ignore them.

Training uses our historical backtest table: for every past day we know what the price actually did seven and thirty days later, so the model learns from real realized outcomes rather than from human-labeled intuition.

Latest test-set accuracy

  • 64% directional accuracy on the 7-day horizon
  • 80% directional accuracy on the 30-day horizon

Accuracy reported on a held-out 20 percent test split from the most recent training run. A completely random three-class guess would score about 33 percent, so both models are learning real signal. The 30-day horizon scores higher because longer-term trends are noisier in the short term but cleaner over time.

The model is retrained roughly once per week against the latest data. You can toggle between the baseline and ML projections on the predictions page using the "Model" pill.

What "confidence" actually means

Confidence is the model's own honesty score, displayed as a percentage from 0 to 100. It combines three things:

  • How much price history we have for that variant, relative to the lookback window the horizon needs.
  • How volatile recent prices have been. A choppy series gets a confidence penalty because chop is harder to forecast.
  • How many of the underlying signals we were able to compute. Missing signals lower the score.

Cards with a confidence below 30 percent show a "Data limited" badge so you can spot them at a glance. The confidence slider on the predictions page lets you filter to higher-conviction rows only.

What predictions do and do not tell you

Predictions answer

  • Which direction is the price likely to move over the next 7 or 30 days?
  • Roughly how big could that move be?
  • How confident is the model in this read, given the data it has?

Predictions do not answer

  • The exact day or hour to buy or sell.
  • The impact of news, reprints, ban list changes, or surprise set announcements.
  • Whether you personally should make a trade. Predictions are not financial advice.

How accurate are the predictions, really?

Accuracy is something you should be able to verify, not take on trust. Every prediction we make is replayed historically against the realized price change, and the results inform the confidence score and the calibration of the magnitude estimate. Hit rate is tracked per horizon, per confidence band, and per signal type, and we publish honest accuracy figures alongside the model description below.

Accuracy is not static. The models retrain against fresh data weekly, and the price history they learn from grows by one day every day, so the longer WealthFamePower runs the better the predictions get.

Disclaimer. Price predictions on WealthFamePower are statistical projections based on historical TCGPlayer market data and other public signals. They are provided for informational and educational purposes only and are not investment, financial, or trading advice. Past performance does not guarantee future results. You are responsible for your own buying and selling decisions.

Frequently asked questions

How often are price predictions updated?
Predictions are recomputed once per day, right after the daily TCGPlayer market price sync that runs at 4:00 PM Eastern. Each day's predictions reflect the most recent closing market price for every tracked variant.
Are these predictions financial advice?
No. Price predictions are statistical projections based on historical TCGPlayer market data and should be treated as one signal among many, not as buy or sell recommendations. Past price behavior does not guarantee future results, and trading card prices can react to news, reprints, tournament results, and other events that the model does not see.
What is the difference between the baseline model and the ML model?
The baseline model is a transparent, rule-based statistical projection built from three signals: momentum, mean-reversion, and volatility. It runs automatically every day. The ML model is a LightGBM gradient-boosted ensemble trained on roughly 30 features per card -- price lags, rolling stats, returns, card metadata, social mentions, and the baseline signals themselves -- and it produces direction and magnitude predictions for every mapped card. The ML model is currently in beta and is re-trained weekly.
Why does my card not appear in the predictions list?
A card is only included in predictions if it is an English-language card with at least some recent TCGPlayer market price history and a confirmed mapping to a TCGPlayer product. Japanese-only cards, sealed products, and cards without enough price history are excluded. If the prediction has very low confidence the row may also be filtered out by the default confidence slider on the predictions page.
How is the confidence percentage calculated?
Confidence combines three things: how much price history we have for that card relative to the horizon window, how volatile its recent prices have been (volatile series get a lower confidence), and how many of the underlying signals we were able to compute. A confidence near 100 percent means we had a full lookback window of stable prices and all signals were available. A confidence below 30 percent surfaces the 'Data limited' badge on the card tile.
How far back does the price data go?
We started accumulating daily TCGPlayer market price snapshots in late January 2026 and have been collecting daily ever since. As of writing, that gives the models roughly 111 days of contiguous history to work from. The price history grows by one day every day, and prediction accuracy improves alongside it.