btc scorecard

Loading…

Calibration of the one-day interval

Each bar is the share of scored days whose outcome fell in that tenth of the predictive distribution (the probability integral transform). A calibrated model fills the ten bins evenly; a hump in the middle means the intervals are too wide, piles at the ends mean too narrow, a slope means a biased centre.

Model against baselines

Log score = log predictive density at the outcome; higher is better. The model is scored against the same forecast with zero drift (does the +0.1 %/day centre earn its keep?), against a random walk with the trailing 30-day realized volatility (does the volatility model — weekday, annual, lunar terms, mean-reverting level — beat the naive one?), and against a centre that adds the four-year halving-cycle harmonics (a lead from the historical walk-forward at the one-year horizon, shown on the members’ chart). Cumulative differences and a sign test, by horizon.

Pre-registered leads from the two rounds of study

Scored on days after the fitting sample only (from ), in the direction fixed before those days existed. At the in-sample effect sizes these need years, not weeks; the tallies are shown so that nobody has to take that on faith.

Recent one-day predictions

Each line was written at the close of the anchor day and scored at the close of the target day; nothing is revised. Open rows are awaiting their close. A dot after the realized price marks an outcome inside the 50 % band — about half of them should carry one.