GARCH volatility¶
Conditional volatility as a causal feature, from a GARCH(1,1) fit. The
parameters are estimated by maximum likelihood on a training prefix (optionally
refit on the expanding window every refit steps), then the conditional
volatility \(\sigma_t\) is forward-filtered over the whole series — which is
causal because \(\sigma_t\) is \(\mathcal F_{t-1}\)-measurable. The first
min_train values (the in-sample warmup) are returned as NaN.
The single authoritative ARMA/GARCH implementation lives in
fynance.models.econometric_models (the Numba recursion) and
fynance.estimator (the likelihood); this feature only wires the thin
fit + forward-filter on top.
|
Causal GARCH(1,1) conditional-volatility feature. |