# garch_regime / 1

Hypothesis. Volatility clusters and is forecastable; the days that follow a high conditional-volatility forecast carry more downside than upside for a long-only holder. Staying out of the market when a GARCH(1,1) model announces a turbulent day, and following the daily trend otherwise, should cut drawdowns more than returns.

Rule. Hourly bars are reduced to the last close of each complete UTC day (180 days of lookback). For each symbol: (1) trend: last daily close above the mean of the last 30 daily closes; (2) regime: fit a zero-mean GARCH(1,1) (arch) on the daily log-returns of the window and compare the one-day-ahead volatility forecast with the median fitted volatility of the window; the regime is calm if forecast <= 1.25 x median. Weight 1/N when both hold, otherwise USDC. If the optimiser does not converge the symbol is not held.

Cadence. The output only changes when a UTC day completes, and a fit costs tens of milliseconds, so the backtests of this strategy use --decision-every 24 (one decision per day, at 00:00 UTC). The live engine should call it at the same cadence.

What would refute it. GARCH forecasts react AFTER a large move (the shock enters the variance with a lag of one day), so the filter may step aside after the fall and miss the rebound. About 180 daily observations is a short sample for a GARCH fit: the parameters are noisy and the threshold 1.25 is arbitrary.