# Forecast experiment protocol

Recorded before fitting the 200WMA forecasting models. Cutoff: the completed
Friday week ending 2026-10-02. No tactical KISS indicators enter this experiment.

- Use observed daily prices, take each completed week's last available close,
  require consecutive weekly observations, and compute a trailing 200-close mean.
  Do not manufacture weekly observations from monthly prices. Start gold in the
  floating-price era (1971-08-16) when long history is available.
- Features: price/200WMA, its natural log, annualized 13/26/52-week 200WMA growth,
  and annualized least-squares log-200WMA slope over 52 weekly observations.
  The primary specification uses 52-week growth; the others are sensitivity checks.
- Targets: subsequent 1/3/5-year CAGR for BTC; also 10 years for gold. Match to
  the first observed weekly close at or after the calendar anniversary and
  annualize using actual elapsed days/365.25. No extrapolated target endpoints.
- Fit expanding-window OLS to CAGR, using only targets whose endpoint strictly
  predates the forecast origin. Require 104 matured training observations.
  Predict at every eligible weekly origin. This minimum is not 104 independent
  long-horizon returns; report non-overlapping effective sample sizes separately.
- Compare a constant 20% BTC / 7% gold prior, slope-only, extension-only,
  slope+extension, and a fixed 50/50 prior/combined forecast. Do not tune blend
  weights. Also test terminal multiples using the historical median extension
  available at the forecast date, with either 52-week WMA growth or the prior.
- Use common OOS origins for comparisons. Report MAE, RMSE, sign accuracy,
  correlation, OOS R² relative to the fixed prior and expanding historical mean,
  subperiod results, and non-overlapping offset checks. Do not clip bad forecasts
  out of forecast error metrics. Report nonphysical predictions separately.
- Treat overlapping forward returns as dependent. Use horizon-length block
  bootstrap intervals for loss differences and HAC descriptive coefficient
  uncertainty. Where fewer than three full OOS horizon blocks exist, explicitly
  mark statistical evidence insufficient. Do not present thousands of overlapping
  weeks as thousands of independent experiments.
- Economic diagnostic: a predeclared γ=3 single-risky-asset/cash Merton problem
  with fixed volatility (BTC 60%, gold 20%) and 3% cash CAGR. Compare allocations
  and next-week portfolio returns on the same forecast origins. This isolates
  forecast usefulness; it is not an investable multi-asset v2 backtest.
- Use observable price history vintages available now, disclose revisions and
  early-market quality. A 2026 selected prior is a retrospective benchmark, not
  evidence that an investor specified it in 2014.
- Promote a 200WMA estimator only if OOS gains are economically meaningful,
  stable across time/slopes, and supported by enough independent outcomes.
  Otherwise retain explicit structural scenarios with broad uncertainty.

Production selection remains a research judgment; no single in-sample statistic
or hindsight-winning variant is sufficient. Keep all tested variants in results.

Implementation correction after fitting: nonoverlapping offsets and horizon-block counts use ceil(365.25 × H / 7) weekly observations, rather than 52 × H, to avoid a one-week overlap with calendar-anniversary labels. Forecasts and model selection are unchanged.