Research track v2 — state-machine variant
Evaluation window 2011-01-03 → 2026-09-17 (ETF era only) · generated 2026-09-17T16:46:15+00:00 · Live dashboard · v1 backtest
Phase 1
Tactical warning
Phase 2
≥15% drawdown episode (numbered)
SPY adjusted close (log scale)
Side-by-side: current system vs v2
| Metric | Current system 2011–2020 (in-sample) | v2 2011–2020 (in-sample) | Current system 2021–2026 (out-of-sample) | v2 2021–2026 (out-of-sample) |
|---|---|---|---|---|
| Episodes (≥15%) | 3 | 3 | 2 | 2 |
| Hit rate (warned at peak) | 0 / 3 (0%) | 0 / 3 (0%) | 0 / 2 (0%) | 0 / 2 (0%) |
| Median lead (hits) | n/a | n/a | n/a | n/a |
| False-positive share of red days | 0.0% (0 / 16) | 0.0% (0 / 45) | 0.0% (0 / 7) | 85.5% (289 / 338) |
| State flips | 2 | 2 | 12 | 8 |
Episodes in the evaluation window
| # | Peak | Trough | Max drawdown | Current at peak | v2 at peak | v2 lead / lag | v2 red share of decline | Sample |
|---|---|---|---|---|---|---|---|---|
| 1 | 2011-04-29 | 2011-10-03 | -18.6% | PHASE 1 | PHASE 1 | no warning | 0% | in-sample |
| 2 | 2018-09-20 | 2018-12-24 | -19.3% | PHASE 1 | PHASE 1 | no warning | 0% | in-sample |
| 3 | 2020-02-19 | 2020-03-23 | -33.7% | PHASE 1 | PHASE 1 | -20 td (lag) | 17% | in-sample |
| 4 | 2022-01-03 | 2022-10-12 | -24.5% | PHASE 1 | PHASE 1 | no warning | 0% | out-of-sample |
| 5 | 2025-02-19 | 2025-04-08 | -18.8% | PHASE 1 | PHASE 1 | -34 td (lag) | 3% | out-of-sample |
v2 mechanics
- Entry: PHASE 2 opens when a Tier 1 gate (credit or yield) has fired within the trailing 60 trading days AND at least 2 of the 4 Tier 2 confirmations have fired within the trailing 15 trading days. TACTICAL WARNING is active whenever the yield gate has fired within the trailing 60 days without a completed quorum.
- Exit: an open PHASE 2 holds until the explicit exit rule fires — DGS10 ≤ 4.00% AND credit spread < 320 bps — which also resets any amber state to green. Nothing else closes the state (this hysteresis is the point of the variant: fewer flips, no flickering).
- Yield gate: from 2022 onward, the absolute 5.10% threshold; before 2022, regime-relative — DGS10 at or above the 90%th percentile of its trailing 756 trading days (~3 years), shifted one day so the window never includes the current day.
- Credit gate and data break: identical thresholds to the live system (375/425 bps). As documented in the v1 backtest, FRED serves only a rolling 3-year window of ICE BofA history, so before ~2023-09 the gate and the exit rule's credit leg run on the Moody's BAA10Y proxy (tighter than HY OAS — pre-2023 readings are conservative), and the single-B leg is inactive.
Sample discipline
- All parameters (the 60/15-day latches, the 4.00% / 320 bps exit rule, the 3-year p90 window) were fixed against the 2011–2020 in-sample period only. 2021–2026 is reported as untouched out-of-sample: nothing was calibrated, tuned, or selected on it.
- ETF availability: GLD starts 2004, CPER and BNO 2010, SPHB/SPLV May 2011, and the SPHB/SPLV leg additionally needs a 200-day window, contributing only from ~early 2012. ETF-based indicators do not exist before 2011, so this track evaluates 2011-01-01 onward and makes no claims about pre-2011 performance.
- Metrics use the v1 definitions: hit = amber/red on the peak day; lead = trading days from the start of the non-green run containing the peak; false positive = PHASE 2 day within 5% of the running all-time high with no ≥15% episode beginning in the next 126 trading days; flips = days where the state differs from the previous day.
Limitations
- This is a research track, not the live system; the live monitor is unchanged. Parameters are exogenous design choices — only the sample-split discipline is claimed, not optimality.
- Fewer state flips are guaranteed by construction (hysteresis); that is not by itself evidence of skill.
- The exit rule is asymmetric by design: in a high-rate regime (DGS10 above 4.00%) PHASE 2 cannot close even after credit normalizes, so the system can stay red long after the risk has passed. The out-of-sample false-positive share shows that cost directly.
- The pre-2022 regime-relative yield trigger is trailing-only, but its window and percentile were still chosen with in-sample knowledge.
- Same data caveats as the v1 backtest: FRED publication lags ignored, no vintage adjustment, monthly copper proxy forward-filled, closing prices only.