← All research|Also available in 中文

Regime Autopsy — does the Shield survive its worst regimes?

Published: 2026-08-07

Simulated results — no real money. Every figure below is computed from historical price data by a backtest. No capital was invested and no orders were placed. Backtests apply a strategy to the past with full knowledge of how that period turned out, and the figures exclude slippage and liquidity effects, though a simulated 0.1% exchange fee is charged on every rebalance, DCA buy and redeployment. Simulated and past performance is not a reliable indicator of future results. AQMath is software, not investment advice.

Date: 2026-08-07
Engine: Aegis (v14) Dual-Speed E2E — identical wiring to the live paper trading service (180-day KKT risk-parity MACRO loop with the ADV-K2 liveness screen + Aegis (v14) Deleverage Shield)
Status: 📊 RESEARCH — worst-regime out-of-sample stress study, no basket promoted

1. Objective

Every published study so far scored the Shield on whole windows. The question left open is the one critics ask first: what happens inside the worst segment — a crash regime the optimizer never saw during fitting?

This study slices the frozen-weight walk-forward series post-hoc along six named crash regimes and measures the Shield's max drawdown against Buy & Hold inside each one. Every regime is scored by weights and parameters that were frozen before it happened — by construction out of sample.

No parameters were changed. Every run uses production default_config() and the unmodified production code path — walk-forward, no lookahead. This is also the first study to run with the ADV-K2 liveness screen active at every re-optimisation, exactly as in the production service.

2. Baskets and data

Three structurally different long-history baskets, each anchored with PAXG:

Six regime windows were cut on every basket: R1 Mar-2020 crash, R2 May-2021 unwind, R3 LUNA contagion, R4 FTX collapse, R5 Aug-2024 carry unwind, R6 2025-26 corrections.

Settings: start $1,000, DCA $100 every 30 days, 10 bps simulated fee on every trade, 180-day optimizer warm-up, MACRO re-optimisation every 180 days (12-13 re-optimisations per basket).

Data: Yahoo Finance daily closes (unadjusted), fetched 2026-08-07; common-date alignment per basket. This study uses a different data source than the earlier CoinGecko-based studies; daily-close figures are comparable but not bit-identical.

3. Headline result — per-regime drawdown cuts

Max drawdown inside each regime window, Shield vs Buy & Hold. TWR daily returns, DCA flows removed, rf 5%.

15 of 15 covered regime-basket combinations show a positive drawdown cut. The worst segments: LUNA contagion held to 16.8% vs 44.1% on basket A and 21.3% vs 44.8% on basket B; FTX collapse cut to 3.3% vs 13.3% (A) and 1.9% vs 9.2% (B).

R1 (Mar-2020 crash) is NOT covered — the PAXG anchor plus the 180-day warm-up starts baskets A and C on 2020-03-24, after the flush bottom. Reported as a gap, not hidden.

Note on basket C: its Buy & Hold column includes the dead CEL token for the whole window — that is the point of the basket. The Shield side runs with the liveness screen, which zeroed CEL's weight once (2025-08-24, see the liveness screen study).

4. Full-window runs

Each basket on its full walk-forward window (warm-up included).

Reading: the signature is the same as in every earlier study — the Shield gives up absolute return in windows that contain strong bull legs and buys it back in drawdown. On baskets A and C the full-window risk-adjusted efficiency improves (Calmar 0.57 vs 0.38 and 0.65 vs 0.50). Basket B is the honest exception: it was the weakest basket of the three, and on B the full-window Sharpe and Calmar favour Buy & Hold — the Shield's case on B rests entirely on the per-regime drawdown cuts above, not on full-window efficiency.

5. Visualizations

Virtual equity, all three baskets on one absolute-time axis (log scale):

Drawdown — Shield vs Buy & Hold, full histories:

Per-regime max-drawdown cut (percentage points), grouped by basket:

6. Gap vs drift — what a daily-close system can catch

The Shield reacts to daily closes. A same-day open gap is uncatchable by any close-based system. Decomposition of each regime's drawdown on the equal-weight risky basket (no PAXG) into the close-to-open leg (gap) and the open-to-close leg (drift):

Crypto crashes are drift-dominated: 95-100% of each regime's damage unfolds in the seen, intraday leg that a daily-close system can react to. The uncatchable residual is concentrated in single worst days (May 19 2021: −30% on the equal-weight risky basket).

7. Caveats (disclosed, not hidden)

• R1 Mar-2020 is uncovered by design (PAXG anchor + 180-day warm-up start the baskets on 2020-03-24). The study cannot speak to that crash.
• R6 is defensive-heavy. In the 2025-26 window the Shield spent roughly 40% of days parked in USDC (236 of 583 regime days on basket B). That is the same defensive bias visible in the live forward log — it protects, but it also costs participation.
• This article replaces a withheld earlier draft. The first version of this study ran without the liveness screen, and the volatility cap gave the dead CEL token a 24.7% frozen weight in one basket-C re-optimisation, contaminating the results. That flaw is fixed in production and in this study — see the liveness screen article. The screen fired exactly once across all 38 re-optimisations (CEL, 2025-08-24); baskets A and B are bit-identical to their unscreened runs.
• Data source change: Yahoo Finance (unadjusted closes) instead of CoinGecko; figures are comparable to, not identical with, earlier studies.

8. Methodology

1. MACRO loop: KKT risk-parity optimisation on the trailing 180-day window, re-run every 180 days; weights frozen in between (12-13 re-optimisations per basket). After each re-optimisation the ADV-K2 liveness screen zeroes the weight of any token that failed the $1M/day trailing-volume test at two consecutive re-opts (fail-open, trailing data only) — the same screen running in the production service. 2. DAILY loop: Aegis (v14) Deleverage Shield evaluated on each close; threshold rebalancing only trades when the target drifts beyond the deadband; 10 bps fee on every trade, DCA buy and redeploy. 3. Regime cuts: post-hoc slices of the frozen-weight walk-forward series along named crash windows — each regime is scored by parameters frozen before it. Minimum 20 days inside a slice to be scored. 4. Metrics: TWR daily returns (DCA flows removed), rf 5%; max drawdown from the virtual equity peak. Buy & Hold always tracks the full raw basket composition, including tokens the screen zeroes. 5. Data: Yahoo Finance daily closes (unadjusted), fetched 2026-08-07; OHLC for the gap/drift decomposition from the same source. 6. Code path: paper_trading.daily_step / macro_reoptimize imported unmodified from the production service — bit-for-bit the math that runs the public forward log.

9. Verdict

• 15/15 covered regime-basket combinations show a positive out-of-sample drawdown cut. The modulator holds on regimes it never saw.
• Headline worst segments: LUNA contagion 16.8% vs 44.1% (A) and 21.3% vs 44.8% (B); FTX collapse 3.3% vs 13.3% (A).
• The return cost is real and visible in every CAGR column: the Shield trails Buy & Hold on absolute return in all three full windows.
• Basket B shows the honest limit of the claim: per-regime cuts yes, but full-window risk-adjusted metrics favour Buy & Hold on that basket.
• With the liveness screen active the landmine basket is clean — the dead token never carries real weight after 2025-08, and the full-window Calmar there is the best of the three (0.65).
• No basket promoted. Recorded as research evidence only.

Simulated results — no real money. Every figure is computed from historical price data by a backtest; no capital was invested and no orders were placed. Figures exclude slippage and liquidity effects, though a simulated 0.1% fee is charged on every trade. Simulated and past performance is not a reliable indicator of future results. AQMath is software, not investment advice.