Historical validation

Methodology

• The exact thresholds and state rules from monitor.py are re-applied to each trading day using only data available up to that day (trailing windows only; no centered or forward-looking values in the signal itself). FRED series are aligned to SPY's calendar by last-observation-carried-forward.
• Where an ETF did not yet exist, its price leg is spliced to the FRED proxy above; the proxy history is back-adjusted to the ETF's price level at the seam so the series is continuous. Spliced legs are listed and marked in the table above.
• An episode opens at a peak (highest close of the trailing 252 trading days) once the close falls ≥15% below it, and ends at the trough; scanning resumes at recovery to the peak level or at the next 252-day high, whichever comes first. This keeps secular bears (2000–02, the GFC) as single episodes while still catching independent declines that never set an all-time high (2010, 2011). "Lead" counts trading days from the start of the amber/red run containing the peak; negative means the first signal came after the peak. An episode is a "hit" when the signal was amber or red on the peak day.
• A false positive is a PHASE 2 day with SPY within 5% of its all-time high and no ≥15% drawdown episode beginning within the following 126 trading days (~6 months). The forward-looking part is used only to score history; it is never part of the signal.
• Summary metrics: hit rate = hits / episodes; median lead over hit episodes; false-positive share = false-positive days / all PHASE 2 days (353 of 7,475 days were red).

Limitations (read before trusting this)

• ETF history is short: SPHB/SPLV start May 2011, BNO mid-2010, CPER late 2010, GLD late 2004. Before the splices, proxies stand in; before ~2012 the SPHB/SPLV leg does not exist at all, so the quorum runs on 3 of 4 confirmations and PHASE 2 is harder to reach — the early record is therefore conservative, not directly comparable to the modern record.
• Credit gate data break: since September 2023, FRED serves only a rolling 3-year window of ICE BofA index history (licensing), and ALFRED vintage queries are blocked for these series. Before the ICE window (~2023-09), the credit gate runs on the Moody's Baa–10Y spread with the threshold unchanged, and the single-B leg is inactive. Baa is investment-grade debt — its spread typically runs 100–200 bps tighter than HY OAS — so the gate is structurally harder to trip before 2023. The pre-2023 record is therefore conservative; only the last ~3 years use the live series end to end.
• The gold leg uses COMEX front-month futures (GC=F, from 2000-08) because FRED removed the London gold fixing series in 2020; before 2000-08 the gold/copper leg is inactive, so the pre-2000 quorum runs on only 2 of 4 confirmations (claims and oil). The copper proxy is monthly (forward-filled), so the gold/copper trigger is coarse before CPER.
• FRED publication lags are ignored: series are treated as known on their observation date. In reality initial claims are released about a week late and credit spreads settle a day late, which slightly flatters the reconstructed signal.
• No vintage / point-in-time adjustment: revised FRED history is used as it stands today.
• The 5.10% yield gate is calibrated to the post-2022 rate regime. For most of 2000–2020 the 10Y never reached 5.10%, so Gate 2 was structurally inert for most of this sample; the TACTICAL WARNING state barely appears before 2022. Conversely, in the high-rate 1990s the gate stayed on for months at a stretch — the pre-2000 "hits" and the long median lead are regime artifacts (a permanent warning eventually brackets any peak), not predictive skill.
• 2015–16 is the archetypal false positive for the LIVE monitor: oil/EM credit stress pushed HY OAS through 375 bps with confirmations firing, yet SPY's decline stopped around −13–14%, short of the −15% episode definition. In this reconstruction the Baa proxy peaked at 363 bps, so the gate stays dark and those red days do not appear — the false-positive share here is very likely lower than the live signal's.
• Episode, lead and false-positive statistics are computed on closing prices; intraday path risk is not modeled.