Model Scorecard

Forecasting research · BTC · ETH · XRP · LINK

How well each forecasting model predicted whether BTC, ETH, XRP and LINK would end above or below price targets 30 to 90 days out, scored against what actually happened and compared with the simple base rate.

Bottom line

Today's 30-day call for each coin (UP, DOWN or FLAT), from the model with the best record here, is on the Crypto Price Odds page.

Base rateHow often a move of that size happened in the past. The simplest possible forecast, and the bar every model must clear.

SkillReduction in forecast error (Brier score) versus the base rate. 0% = no better, +10% = 10% less error, negative = worse.

Sealed holdoutThe most recent stretch of history, hidden while models were built and then scored once. The only fully fair test.

P vs QP = real-world forecasts from history and ML. Q = implied by option prices, which include a risk premium, so shown for reference only.

Final holdout verdict

Leaderboard: every model on one asset and horizon

Brier skill compares each model's probabilities with the outcomes, relative to the past event frequency on the same forecasts. 0% means no better than the base rate; positive is better; negative is worse. Bars show the 95% interval.

Period

Outcomes

Asset

Horizon

Model family

Calibration: do the probabilities mean what they say?

When a model says 20%, does the event happen about 20% of the time? Forecasts are grouped into bins; points on the dashed diagonal are perfectly calibrated. Uses the period, outcomes, asset and horizon selected above.

Model

Dot area: number of forecasts in the bin (overlapping windows, not independent)Vertical bar: 95% interval for the observed frequency

Show calibration as a table

Bottom line

Final holdout verdict

Leaderboard: every model on one asset and horizon

Calibration: do the probabilities mean what they say?

How this was tested