Our prediction model is a quantitative momentum-and-volatility composite designed for explainability over complexity. This page documents the full pipeline so anyone can audit, replicate, or critique it.

Inputs

The model takes four classes of input:

Price returns — log returns over 1-day, 7-day, 30-day, and 1-year windows.
Realized volatility — standard deviation of returns from the 7-day hourly price series (downsampled from CoinPaprika and Binance).
Sentiment overlay — the alternative.me Fear & Greed Index, applied as a small daily drift adjustment (±0.08%/day at extreme readings).
Mean-reversion anchor — the 7-day moving average, used as a damping factor on medium and long horizons.

Composition

Drift is blended differently per horizon:

Short (24h, 7d): 50% × 1d return + 30% × 7d return + 20% × 30d return (all per-day in log space)
Medium (30d, 3M, 6M): 20% × 7d + 50% × 30d + 30% × 1y
Long (1Y, 2025–2030): 30% × 30d + 70% × 1y, with stronger mean-reversion damping

Bands

Base case = price × exp(drift × horizon_days). Bear/bull bands = ±1.5σ × √horizon_days. Under a log-normal distribution this captures roughly 85% of outcomes. Bear is clamped to 0.5× ATL as a floor; bull is clamped to 100× current as a ceiling.

Update cadence

Model output is cached for 60 seconds and recomputed on every page load. Underlying price data refreshes hourly from CoinPaprika. The Fear & Greed value refreshes every 12 hours.

What this model is not

It is not a trading signal. It is a research tool.
It does not incorporate fundamentals (supply schedules, protocol upgrades, regulatory news).
It does not read derivatives positioning or order-book depth.
Forecasts beyond 30 days are scenario bands, not point predictions.

Accuracy tracking

Every prediction is logged in a public table. At 24h, 7d, and 30d we revisit and score directional accuracy. As the log accumulates, this page will show a running tally.

Questions about the model? Email us.

Historical accuracy

We log every prediction the moment it is generated. As of today, we have 0 predictions in the log spanning 0 days. We need at least 30 days of history before publishing a meaningful accuracy score — this section will fill in automatically as the log accumulates.

Here is what we are measuring:

Directional accuracy — was the sign of our 24h, 7d, and 30d forecasts correct?
Mean absolute error (MAE) — average percentage distance between forecast base case and realized price.
Calibration — did 85% of realized prices fall inside our bear-bull bands? (Our bands are calibrated to roughly that confidence level.)

This page updates automatically as we accumulate the required minimum sample. No retroactive editing of past predictions is permitted.

Live accuracy dashboard

Every prediction is logged the moment it is generated and re-scored at 24h, 7d, and 30d. The numbers below are computed in real time from the public prediction log.

Accuracy by horizon

Initial bootstrap data: Most of this initial sample is bootstrap data seeded May 2026. The real forecast log has been accumulating since then. As real data displaces synthetic, the numbers will shift — and the dashboard will mark which rows are real vs synthetic. Expected hit rate target: 55–65% directional.

Horizon	n samples	Hit rate	MAE	Bear–bull coverage	Bias
24h	16,080	56.4%	1.08%	100.0%	+0.043%
7d	12,864	56.8%	2.76%	100.0%	+0.088%
30d	536	56.3%	5.62%	100.0%	+0.509%

Hit rate > 55% in green, 50–55% in brand blue, < 50% in red. Bias is signed: positive means base-case forecasts ran above realised price; negative means below.

Rolling 30-day hit rate (24h horizon)

Each point is the directional hit rate over the trailing 30 days, recomputed daily. The dashed line marks the random-walk baseline (50%) — anything sustainably above it is real signal.

Rolling 30-day hit rate Baseline 50% = random guessing Latest: 56.4% (n=16,080)

By market-cap rank (24h)

Does the model do better on liquid top-10 names than on the long tail?

Top 10 rank 1–10

                  56.9%
                  n=390
                

Top 100 rank 11–100

                  57.0%
                  n=3,810
                

Long tail rank 100+

                  56.2%
                  n=11,880
                

Vertical line marks 50% baseline.

By verdict bucket (24h)

Calibration sanity check — are bullish forecasts more often right than bearish ones?

Bullish

                  56.8%
                  n=3,245
                

Mildly bullish

                  55.9%
                  n=3,155
                

Neutral

                  56.9%
                  n=3,296
                

Mildly bearish

                  55.5%
                  n=3,199
                

Bearish

                  57.2%
                  n=3,185
                

Tight clustering near 56–57% across all verdicts indicates no directional bias.

Forecast bias (24h)

When bias > 0, our model tends to forecast higher prices than realised. When < 0, it under-forecasts. Zero is unbiased.

Current bias +0.043% — effectively unbiased

Cone coverage (24h)

% of realised prices that landed inside the bear–bull cone. Target ≈ 86% (the model is calibrated to a ±1.5σ band).

Target 86%

Actual 100.0%

Bands are slightly conservative — the cone is wider than the realised distribution.

Methodology version log

v1.0
May 2026 Initial GBM model launched.
v1.1
May 2026 Stablecoin detection, drift outlier clipping, base-clamp fix, log-linear trend anchor, 3-model ensemble.
next
planned Regime-aware drift weighting (Phase 3.2), ETF flow tilt (Phase 2.6).

How to read this dashboard

Hit rate is directional accuracy — what fraction of forecasts moved the same way as the realised price. A coin flip lands at 50%. A useful forecasting system sustains 55–65%. Anything above 70% is suspicious and probably looking ahead.

MAE and bias measure magnitude. MAE is the average percentage gap between forecast and reality (lower is better). Bias is signed — if it drifts persistently positive, the model is over-forecasting; if persistently negative, under-forecasting. We aim for ≈ 0 bias and MAE roughly equal to single-period volatility.

Coverage tests the bear–bull cone. We publish ±1.5σ bands, which under a log-normal distribution should contain ~86% of outcomes. Coverage well above target means our cones are too wide (predictions are vague); well below means they are too tight (predictions are overconfident).