H-028: A daily model gives every coin's chance of a bigger move than usual better than its base rate
Pass Family: forecast · Status: passed
This hypothesis passed its pre-registered test. That is a reason for independent confirmation and paper trading, not for trading real money.
Can a daily model predict big crypto moves? This idea met the bar that was set before testing. It deserves a closer, independent look. It is not a promise of profit.
Result
pooled Brier skill +0.0383 (p = 0.000, need p < 0.025); skill above zero on 13/13 coins
| Asset | Days | Brier skill | p-value | q (FDR) | Guessed right |
|---|---|---|---|---|---|
| BTC | 1760 | +0.0172 | 0.001 | 0.001 | 56.3% |
| ETH | 1760 | +0.0282 | 0.001 | 0.001 | 56.4% |
| SOL | 1760 | +0.0246 | 0.001 | 0.001 | 57.3% |
| BNB | 1760 | +0.0300 | 0.001 | 0.001 | 58.1% |
| XRP | 1760 | +0.0406 | 0.001 | 0.001 | 57.5% |
| ADA | 1760 | +0.0439 | 0.001 | 0.001 | 57.6% |
| LTC | 1760 | +0.0152 | 0.001 | 0.001 | 56.4% |
| LINK | 1760 | +0.0372 | 0.001 | 0.001 | 57.8% |
| DOGE | 1760 | +0.0542 | 0.001 | 0.001 | 57.4% |
| XLM | 1760 | +0.0564 | 0.001 | 0.001 | 59.0% |
| TRX | 1760 | +0.0364 | 0.001 | 0.001 | 59.5% |
| EVER | 1306 | +0.0715 | 0.001 | 0.001 | 58.7% |
| GRAM | 946 | +0.0600 | 0.001 | 0.001 | 57.7% |
| ∑ | 21612 | +0.0383 | 0.001 | — | 57.7% |
Calibration: forecast vs what happened
| Forecast | Forecast ⌀ | Happened | Days |
|---|---|---|---|
| 10%–20% | 18.6% | 44.4% | 9 |
| 20%–30% | 27.1% | 33.3% | 615 |
| 30%–40% | 36.1% | 38.1% | 4426 |
| 40%–50% | 45.1% | 43.8% | 8013 |
| 50%–60% | 54.4% | 51.8% | 5819 |
| 60%–70% | 64.0% | 62.2% | 2079 |
| 70%–80% | 73.9% | 73.9% | 547 |
| 80%–90% | 83.6% | 83.8% | 99 |
| 90%–100% | 92.8% | 80.0% | 5 |
Why it might work
Same model and inputs as H-027, different question: the probability that the next day's absolute log move is larger than the coin's median absolute daily move over the past 365 days (measured up to D, so the base rate is near one half). Price direction is close to unpredictable, but the size of moves clusters in calm and nervous stretches, so this is where a model can show real skill; it would help size positions and avoid big losses, not make money by itself. One logistic regression shared by all 13 coins. Inputs at the close of day D: log returns over 1, 3, 7 and 30 days, distance from the 50- and 200-day averages, 7- and 30-day volatility, Bitcoin's 1- and 7-day returns, and, lagged one day for publication delay, BTC funding (7-day mean), Fear & Greed, 30-day USDT+USDC supply growth and 7-day BTC net exchange flow. Every input is a rolling 365-day z-score of its own past (min. 180 days, clipped at ±5), so coins with very different volatility share one model. The model is refitted on the first day of every month on decisions whose outcome was already known (up to two days before), L2 = 1. Costs: 0.1 % per side on Binance, 0.4 % per side on Gate.io coins (0.2 % fee + 0.2 % slippage). Passes on the same pooled Brier-skill test from 2022 on; the circular shift is conservative here, because rotated forecasts still overlap long volatility regimes. Shares alpha with H-027 (0.05 / 2).
Verdict history
- 2026-10-28: Pass (this engine, data to 2026-10-26)
Pre-registered specification
kind: forecast
target: risk
model:
type: pooled_logit
features:
- r1
- r3
- r7
- r30
- d50
- d200
- v7
- v30
- btc_r1
- btc_r7
- funding
- fear_greed
- stable30
- flow7
zscore_window: 365
zscore_min_periods: 180
clip: 5
l2: 1.0
refit: monthly
external_lag: 1
ffill_limit: 3
universe:
- BTCUSDT
- ETHUSDT
- SOLUSDT
- BNBUSDT
- XRPUSDT
- ADAUSDT
- LTCUSDT
- LINKUSDT
- DOGEUSDT
- XLMUSDT
- TRXUSDT
- EVERUSDT
- GRAMUSDT
costs:
default: 0.001
EVERUSDT: 0.004
GRAMUSDT: 0.004
periods:
- name: oos
start: '2022-01-01'
significance:
method: circular_shift
n: 2000
alpha: 0.025
criterion:
metric: brier_skill
pooled: true