`snapshot` writes each tracked model's forecast quantiles at the seven
backtest horizons from the latest price to data/forecasts/<origin>.csv. It
refuses stale data (older than two days) and duplicate dates, so snapshots
can't be reconstructed after the fact; committing them dates them.
`forward` scores every recorded forecast whose target date has passed,
reusing the backtest's scoring (now factored out as evaluate.score).
Tracked: random_walk, drift_rw, cycle, powerlaw, plus powerlaw_ou,
powerlaw_ou_param and cycle_on_powerlaw, which development data couldn't
settle. `just weekly` runs update, snapshot and forward.
First snapshot: 2026-09-23 (BTC $84.4K). The first outcomes are due
2026-10-23.
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.
Models are now a Composite of drift, volatility and (optional) shape
components, so an experiment can swap one part against a fixed control.
btcmodel/experiments.py holds seven experiments built from the ideas in the
old branches (catalogued in docs/2024-ideas.md), each with its hypothesis
and source, and a verdict rule fixed before anything ran. `just ab` runs
them on development data. Results:
- Shrinking the cycle drift, and a power-law trend (plain or reverting),
beat their controls. The power law beats the random walk by 53-63% at
3-4 years with unbiased outcomes, so it is promoted to MODELS.
- Every alternative volatility estimate (EWMA blends, other windows,
reversion to a level or trend) is worse than the trailing 365-day window.
Cycle-dependent volatility, heavy tails and stretched cycle phase show no
reliable effect.
Replace the 2024 model (model.py, ~2000 lines) with the btcmodel package, the
baseline for future work:
- Forecasts are quantiles of log price at each horizon, scored with CRPS in a
walk-forward backtest (origins every 30 days from 2014, horizons 1 month to
4 years). Skill is relative to a zero-drift random walk, with circular
block-bootstrap intervals and a count of independent windows.
- Development data stops at 2024-11-26, the last day the 2024 model saw.
Later outcomes are a holdout, scored only by `backtest --holdout`.
- Models: random_walk, drift_rw, and cycle (the 2024 model's cycle-position
drift, now kernel-smoothed and recency-weighted). On development data
nothing beats the random walk with confidence; cycle loses at every horizon.
- Prices: the Investing.com archive moves to data/ (cut at 2024-11-26; its
last row was intraday) and is extended with Coinbase daily closes by
`update`.
Also: Nix flake dev shell (Python 3.13, pandas 3), ruff in place of black,
pytest suite, and a rewritten README. NOTES.md is removed as inaccurate, and
poetry is dropped.
sam