Round 2: reverting power law and cycle-on-powerlaw tests; one-time holdout run.

Add TrendReversionVol: deviations from the power-law trend follow a daily
AR(1), so uncertainty levels off, optionally plus trend-parameter
uncertainty with an autocorrelation-adjusted effective sample size.

Two experiments, run under the unchanged verdict rule:
- powerlaw-ou: +21% to +45% vs powerlaw at 2-4 years, but slightly negative
  point estimates at 1 month make it inconclusive.
- cycle-on-powerlaw: inconclusive (+18% at 2 years, negative elsewhere).

The holdout (outcomes after 2024-11-26) was scored once, for the four
candidates fixed beforehand. powerlaw is the best long-horizon forecast
(+45% and +58% vs the random walk at 2 and 3 years); nothing beats the
random walk inside a year; cycle fails badly. Results are in the README.

btcmodel/models/volatility.py

@@ -6,8 +6,8 @@ import numpy as np
 import pandas as pd
 
 from ..data import log_returns
-from ..halving import cycle_position
-from .drift import mean_by_cycle_day
+from ..halving import GENESIS, cycle_position
+from .drift import PowerLawDrift, mean_by_cycle_day
 
 
 @dataclass(frozen=True)
@@ -104,3 +104,43 @@ class CycleVol:
         future = history.index[-1] + pd.to_timedelta(np.arange(1, horizons.max() + 1), unit="D")
         _, future_day = cycle_position(future)
         return base * np.sqrt(np.cumsum(ratio[future_day] ** 2)[horizons - 1])
+
+
+@dataclass(frozen=True)
+class TrendReversionVol:
+    """
+    Uncertainty for a price that reverts to the power-law trend.
+
+    Deviations from the trend follow a daily AR(1) with coefficient φ (fitted
+    by PowerLawDrift), so their variance levels off: after h days it is
+    σ²(1 − φ^2h) / (1 − φ²), with σ the trailing `window`-day volatility.
+
+    With `parameter_uncertainty`, the uncertainty of the fitted trend line is
+    added. The residuals are so autocorrelated that ~5000 days carry the
+    information of only n(1 − φ)/(1 + φ) independent points, and the
+    coefficient covariance is inflated to match.
+    """
+
+    window: int = 365
+    parameter_uncertainty: bool = False
+
+    def sd(self, history: pd.DataFrame, horizons: np.ndarray) -> np.ndarray:
+        intercept, slope, phi = PowerLawDrift().fit(history)
+        phi = min(phi, 0.9999)
+        sigma = log_returns(history).iloc[-self.window :].std()
+        variance = sigma**2 * (1 - phi ** (2 * horizons)) / (1 - phi**2)
+        if self.parameter_uncertainty:
+            variance = variance + self._trend_variance(history, horizons, intercept, slope, phi)
+        return np.sqrt(variance)
+
+    @staticmethod
+    def _trend_variance(history, horizons, intercept, slope, phi) -> np.ndarray:
+        t = (history.index - GENESIS).days.to_numpy()
+        x = np.column_stack([np.ones(len(t)), np.log(t)])
+        resid = np.log(history["close"].to_numpy()) - x @ [intercept, slope]
+        n_eff = len(t) * (1 - phi) / (1 + phi)
+        cov = resid.var() * np.linalg.inv(x.T @ x) * len(t) / n_eff
+        # The forecast mean is a(1 − φ^h) + b(ln t_h − φ^h ln t_0) + φ^h ln P_0.
+        decay = phi**horizons
+        g = np.column_stack([1 - decay, np.log(t[-1] + horizons) - decay * np.log(t[-1])])
+        return np.einsum("hi,ij,hj->h", g, cov, g)