Case 3 · Metric definition
Case 3 — Drawdown metric: the circuit-breaker reset the peak
Symptom. The report said "max drawdown: 53.1%". An investor looking at the equity curve could see that it had, at one point, lost almost everything. The two numbers could not both be right.
Root cause. The engine had a circuit-breaker: when drawdown crossed a threshold it halted, and on resume it reset the running peak to the current capital. The drawdown metric was measured against that peak. So the "reported MDD" was only the deepest drawdown within one post-breaker segment — the original, from-the-all-time-high drawdown was never recorded.
auditkit.checks.invariants supports the fix directly: assert the equity
identity every step and treat the peak as an all-time series that is never
modified by a control-flow event.
What the fix changed. Write two drawdown numbers at once:
mdd_reported— segment-local drawdown (what the breaker sees), andmdd_raw— drawdown from the all-time-high capital, never reset.
| Metric | Value |
|---|---|
| Investor-view raw drawdown, full history (fixed) | 98.7% |
| Reported MDD (breaker-local, buggy) | 53.1% |
| Gap | 45.6 percentage points |
Provenance. Both numbers come from the same run — trading_research/run_out/risk_gate_tests.txt,
lev 1.2, 2021: MDD 53.1% | 原始MDD 98.7%. An earlier version of this note
quoted 99.7%, which is a different configuration (lev 1.2, breaker off,
full history, run_out/sens_odds_fix.txt). The conclusion is unchanged; the
pairing was wrong and is corrected here.
Bottom line. The headline number hid a near-total loss. MDD must be dual-written: a control-flow variable may be reset, but the investor-facing metric may not. Reporting only the reset number is, at minimum, materially misleading.
How to reproduce. Any ledger whose cap drops from an all-time high while
a control event resets the internal peak; assert the equity identity and
preserve an unmodified peak series. See cases/README guidance and
tests/test_invariants.py for the invariant form.