Drawdown live estimator

The measurement

Improvement vs production

0.022895

Schema

canli.alphac-drawdown-live-estimator.v1

Cost unit warning. Turnover is a MULTIPLE OF EQUITY per year; a round-trip cost in basis points is a fraction of the NOTIONAL TRADED. The annual drag is turnover * bps/10000 of equity. Dividing by 100 instead overstates it a hundredfold, which has happened once in this project already.

Excluded deliberately. Ledoit-Wolf shrinkage. artifacts/analysis/ledoit_wolf_effective_sample measured it separately: immaterial at the production setting (0.36% ex-ante vol error) and material at a short halflife (4.46% mean, 38% worst). Folding it in here would conflate two effects in one number. It is a PRECONDITION on acting on this result, not a term inside it.

P95 still misses the objective. No configuration holds the 95th percentile at or under 11%. The best is 13.5% at 720/7. The drawdown objective is an EXPECTED maximum drawdown, and both figures are always published together.

Path count caveat. 96 paths, so each expected maximum drawdown carries a standard error of roughly the value reported in max_drawdown_stderr. Configurations share their shocks, so the RANKING is far better determined than any single level, and the ranking is what this analysis is for.

Precondition before acting. At a 21-bar halflife the Ledoit-Wolf shrinkage error measured in artifacts/analysis/ledoit_wolf_effective_sample is 4.46% of ex-ante vol, and at 7 it would be worse — the effective sample shrinks further. That error must be fixed BEFORE the halflife is shortened, not after. And any change to it is a live-book change: owner-gated, and it VOIDS the forward pre-registration draft.

The finding. THE HALFLIFE IS THE LEVER AND THE WINDOW IS NOT. Shortening the halflife inside the production 720-bar window improves expected maximum drawdown monotonically: 10.25% at 720, 9.45% at 63, 8.38% at 21, 7.96% at 7. Shortening the WINDOW buys nothing on top of it — 168/21 gives 8.52% against 720/21's 8.38%, inside the standard error. This corrects a claim published earlier the same day that named the window as the only lever; that came from sampling only halflives at or above 21 days, where the window truncates everything into one narrow band.

Why the original sweep does not transfer. analyze_frontier_14.py updates covariance with an UNTRUNCATED recursion — infinite memory, no window, no seed block. Production's ewma_cov is windowed at cov_window_bars and seeded with an equally weighted block of the oldest cov_min_periods rows, so the window BOUNDS the memory and a halflife longer than the window cannot be delivered. A halflife sweep on an untruncated estimator answers a question about a system this book does not run.

Check it yourself

Every figure above is read from /glassbox/drawdown_live_estimator.json, the artifact the engine wrote. Nothing on this page is typed by hand: the page is generated from that file, so a figure that moves in the artifact moves here and a figure that is not in the artifact cannot appear here at all.

All 21 measurements · The glass box · The research library