How the forecast is made & graded
Each hour the BTC Event Oracle builds a Bitcoin price forecast in two stages — a GJR-GARCH volatility baseline, then a small, hard-capped Claude overlay — and grades every forecast when it matures against a random-walk benchmark using proper scoring rules.
The design goal is honesty, not hype: a transparent quant core, a tightly-bounded AI nudge that can never run away, and an open scorecard that defaults to claiming no skill until the evidence earns it. Here is exactly how each piece works.
1 · The quant baseline
The forecast starts as plain mathematics, before any AI is involved. We model Bitcoin's volatility with a GJR-GARCH process — a variant of GARCH that is asymmetric: it lets downside shocks raise expected volatility more than equally-sized upside shocks, the well-documented “leverage effect.” For crypto returns, which are fat-tailed and cluster in bursts, this asymmetric model fits the data better than simpler constant-volatility or symmetric estimates. A lighter EWMA volatility estimate is kept as a fail-soft fallback if the GARCH fit does not converge.
On top of that, a volatility-regime check classifies the current market as calm, elevated, or turbulent and widens the forecast band when conditions are stressed — so the stated range honestly reflects that the model is less reliable in turbulence. From the resulting volatility term structure we build a lognormal price range and a probability of finishing up (“P(up)”) for each horizon — 1 week, 1 month, and 1 year. The band is always wider for longer horizons, because the further out you look, the blurrier the future.
2 · The bounded AI overlay
Only after the baseline exists does the AI get a turn. Anthropic's Claude reads a condensed set of real-world event signals and applies a small, hard-capped adjustment to the baseline's drift and volatility — never a wild swing. The clamps are deliberately tight (drift moves are limited to a few tens of basis points, the volatility multiplier to a narrow band, and P(up) is pinned away from the extremes), so a confident-sounding model can never hijack the forecast. If the overlay is unavailable or returns anything out of range, the system is fully fail-soft and simply publishes the quant baseline.
The signals Claude reads are each explained, with current readings, on their own pages:
- · Crypto Fear & Greed Index — market sentiment, 0–100.
- · Perpetual funding rate — leverage crowding (longs vs shorts).
- · Open interest — how much leveraged “fuel” is riding on price.
- · Implied volatility (Deribit DVOL) — the options market's forward-looking swing estimate.
- · News tone (GDELT) — whether global Bitcoin coverage is net positive or negative.
3 · How it's graded
Every forecast is timestamped and then scored when it matures against what actually happened — and against the toughest fair benchmark there is, the random walk (assume the price simply stays where it is). We use proper scoring rules, which can't be gamed by shading the numbers:
- Brier score — direction
- Grades the up/down probability. Lower is better; 0.25 is a coin-flip.
- CRPS — full distribution
- Scores the whole predicted range, not just the direction — rewarding sharp ranges that still contain the truth.
- Interval coverage
- How often the real price actually landed inside the stated band, versus how often it should have. The honest check that the ranges are the right width.
- Reliability diagram & PIT histogram
- Calibration receipts: when we say 60%, does it happen ~60% of the time, and are the ranges the right width? Dots on the diagonal and flat PIT bars mean honest numbers.
To decide whether any apparent edge over the random walk is real or just luck, we run a Diebold-Mariano test with Newey-West (HAC) standard errors to handle the overlapping, autocorrelated forecast errors. It defaults to “no significant difference” and only reports skill once the evidence clears that bar — no overclaiming. Finally, we run a head-to-head against real-money prediction markets on the same Bitcoin questions and, once those markets resolve, keep score of who was actually closer.
4 · Why honesty is the whole point
Over short horizons, Bitcoin's price is close to a random walk — tomorrow's move is mostly unpredictable from today's information. So roughly tying the random-walk benchmark is the honest, expected result, not a failure. A site that claimed to beat the market every week would be lying. The value here is the opposite of a crystal ball: a transparent method, well-calibrated uncertainty, and a public record you can audit. See the live numbers, by horizon, on the accuracy page.
This is an educational and portfolio project — model-implied probabilities and ranges, not financial advice and never a guarantee. Bitcoin can lose value fast. Read the plain-English guide →