"Equity-state edge research — final deliverable"

Synthesis of Phases 1–7. Generated artefacts: reports/equity_state_phase{1,2,3,4,5,6,7}.md; ledger reports/equity_state_ledger.jsonl (857 rows); code validation/equity_state/{data,features,phase1,phase2,phase3,phase4,phase5,phase6,phase7,monitor}.py; tests tests/test_equity_state.py.

Promotion recommendation: DO NOT PURSUE (as a tradeable overlay or signal for the existing sleeves). Equity-state variables are genuine, incremental forward-risk forecasters, but they do not improve the frozen sleeves out-of-sample net of the sleeves' own risk controls. No production change was made.

#1. Hypotheses tested

• H1 (Phase 1): market-wide breadth / dispersion / correlation / concentration / leadership carry forward information about market return, market vol, cross-sectional selection, and the frozen sleeves' forward outcomes.
• H2 (Phase 2): conditioning sleeve performance on those states materially changes the forward distribution, incrementally to trailing SPY return/vol.
• H3 (Phase 3): a simple, no-look-ahead state-conditioned exposure overlay improves the frozen sleeves' OOS drawdown/tail without destroying return.
• H4 (Phase 4): (a) the incremental forward-vol signal converts into a state-augmented volatility target that beats a naive vol target; and (b) a sleeve's own lagged equity-curve state predicts its forward return/vol/drawdown incrementally over market state.
• H5 (Phase 5): the state index adds incrementally to the incumbent D1 SPY-200d gate; and cross-sleeve simultaneous deterioration (spec §12) predicts aggregate book risk with a tradeable de-risk overlay.
• H6 (Phase 7): the state index improves the incumbent C7 stress-hedge (the tail/drawdown channel, which does not give up upside the way de-risking does).

#2. Data and universes

• Yahoo total-return _ADJ daily store: 501 current S&P 500 names, 2006-08-14 → 2026-08-12, masked to point-in-time S&P 500 membership (monthly snapshots, step-function; no forward-looking).
• SPY total-return benchmark; GICS sector map (data/cache/sectors_full.json).
• Frozen baselines (unchanged): data/baselines/eqmom_live_v1.csv ("Equity Momentum"), flagship_base_v1.csv ("ML Rotation").
• Limitations: _ADJ holds only current members, so PIT removes late-addition look-ahead but not delisted-name survivorship; mid-caps excluded (2020+ and split-only adjusted); true market cap unavailable, so dollar-volume share is used as a concentration proxy.

#3. Features / method

39 trailing features across five families (breadth, cross-sectional dispersion, correlation, concentration, leadership) and 14 forward targets (SPY return 1/5/21d, SPY vol 5/21d, momentum long-short spread, and the two sleeves' forward 5/21d return, 21d vol, 21d max-drawdown). Methods: rank-IC (full/IS/OOS), effective-n (Newey–West lag 21) p-values, partial IC over trailing SPY return/vol, Benjamini–Hochberg FDR, block-bootstrap CIs / permutation, PIT rolling-rank quintile/tercile partitions, rank-OLS incremental R², and a chronological walk-forward with an untouched holdout (≥ 2023-01-01). Automated future-perturbation look-ahead guards on features and PIT buckets.

#4. Strongest findings

• Forward risk is predictable, incrementally. High concentration (top10_dvwt, hhi_dv) and high cross-market correlation (avg_corr20, corr_spy60) forecast higher forward sleeve volatility and deeper forward drawdowns (BH-significant). Incremental over trailing SPY ret/vol: top10_dvwt → eqmom_vol_fwd21 ΔR² +0.100 (partial IC +0.32); top10_dvwt → eqmom_mdd_fwd21 ΔR² +0.050 (partial −0.22); corr_spy60 → flagship_mdd_fwd21 Q5−Q1 −0.014 (q≈0).
• Joint state matches the hypothesised success condition. The worst forward-drawdown cell is low breadth × high correlation / high concentration / high dispersion (flagship_mdd ≈ −6% vs ≈ −4% elsewhere).
• Correlation → momentum-spread. High correlation predicts a weaker raw momentum long-short spread, sign-stable across all four sub-periods and specific to correlation (dispersion carries no plateau). But the raw spread itself has negative OOS Sharpe and the no-tuning gate fails its placebo (p=0.16).

#5. Failed hypotheses (recorded, not hidden)

• Sleeve return timing: no feature robustly times eqmom/flagship forward returns; strongest OOS ICs ~0.18–0.22 with BH q>0.05 and partial IC ≈ 0.
• Risk overlay (Phase 3): walk-forward A/B/C/D overlays all reduce Sharpe (eqmom best 0.90 vs base 0.97; flagship 0.83 vs 0.84) and give back large return for at best a comparable drawdown; parameter perturbation flips the drawdown effect sign; block-placebo p=0.92 (no better than a randomly permuted state assignment). A single holdout cell (eqmom/dispersion, Sharpe 1.56 vs 1.49, maxDD −21.4% vs −26.0%) is dominated by the multiple-testing and walk-forward failures.
• State-augmented vol targeting (Phase 4): adding the state index to an EWMA vol forecast does not improve a de-risk-only vol-target overlay; state-VT is worse than naive-VT for eqmom and equal for flagship (block-placebo p=0.95 / 0.55).
• Equity-curve feedback (Phase 4, spec §11): no own-curve feature robustly times forward returns incrementally over market state. Surviving pairs are risk-clustering (downside vol → future vol, IC +0.52) and drawdown persistence (flagship deeper drawdown → lower forward return, partial −0.23) — diagnostics, not timing.
• State on top of D1 (Phase 5): the state index improves the rolling walk-forward over D1 (Sharpe 1.01 vs 0.87, block-placebo p=0.00) but the map fitted on all pre-holdout data selected the no-overlay option, so the untouched holdout shows no state benefit — period-concentrated, not robust.
• Cross-sleeve deterioration (Phase 5, spec §12): simultaneous sleeve drawdown predicts higher forward book vol (IC +0.33) but the de-risk overlay loses Sharpe (WF 0.80 vs 0.97; holdout 1.71 vs 1.76) — same de-risk-fails-to-pay pattern.
• State-conditional hedge (Phase 7): hedging in the fragile state looks best OOS 2016-23 (Sharpe 1.13 vs C7 1.05, base 0.98; maxDD −20.4% vs −34.1%) but significantly hurts the untouched holdout (C7+state 1.39 vs C7 1.51, block-bootstrap p=0.02). The apparent hedge skill is entirely the Phase-6 high-vol-regime concentration; in the calm 2023+ holdout the hedge is a drag. Both channels (exposure and hedge) now fail OOS.
• Attribution caveats: avg_corr20 largely proxies trailing SPY vol for vol prediction (partial IC flips sign), and breadth levels predict lower forward market returns (post-rally mean reversion), not a timing edge.

#6. Candidate ranking (best available, for the record)

#6b. Robustness (Phase 6, spec §13D/E/F)

All seven headline relations keep their sign across full / large-cap-100 / two random halves, and rolling 3y windows are sign-stable: breadth→vol 100%, dispersion→vol 99%, correlation→vol 91%, correlation→momentum-spread 96%, breadth→spread 85%.

But two caveats matter for any overlay:

• breadth_sma200 → spy_vol_fwd21 is concentrated in high-vol regimes (IC −0.37 high-vol vs +0.02 low-vol) — the "incremental" breadth info is largely redundant once vol is already elevated, exactly where an overlay would act.
• corr_spy60 → flagship_mdd_fwd21 is regime-unstable (bull +0.09, bear −0.17; 84% rolling-window sign consistency) — the drawdown link is the fragile part of the story.

This is why the risk forecast is robust but the risk overlay could not pay.

#6c. Observe-only monitor (the actionable output)

The only surviving output is a diagnostic. validation/equity_state/monitor.py computes a daily PIT fragility state (breadth / correlation / concentration / dispersion) and writes reports/equity_state_monitor.json + .csv.

• Current reading (2026-08-12): NEUTRAL, state index −0.72, breadth200 0.72, avg_corr 0.009, disp 0.100, top-10 DV weight 0.358.
• Historical state mix: NEUTRAL 42%, PANIC 24%, HEALTHY 18%, FRAGILE 11%, RECOVERY 6%.
• Observe-only: computes no orders, scales no positions, is wired into no bot or the order path. It exists to log "when the book is in a statistically riskier state", not to trade.

#7. Recommendation

DO NOT PURSUE

Equity-state internals are a useful risk monitor / research diagnostic (e.g. an observable "fragile market" flag: low breadth + high correlation + high concentration) but are not a promotable sizing signal for the existing equity sleeves under realistic walk-forward, holdout, cost and placebo testing. Both the exposure channel (Phases 3-5) and the hedge channel (Phase 7) were tested and fail out-of-sample; every apparent gain is concentrated in high-vol regimes and reverses in the untouched holdout. No candidate survives all eight promotion criteria. The only defensible use is the observe-only fragility monitor (validation/equity_state/monitor.py), not an overlay.

Two out-of-scope findings worth separate tickets (no state involved):

1. The incumbent D1 SPY-200d gate is a DD-cut, not an SR uplift, on the current frozen baseline. This was already documented in reports/eqmom_overlay_live_vote.md (D1 OOS SR 1.14 vs base 1.17) — Phase 5 reconfirms it (WF 0.87 vs 0.97; holdout 1.28 vs 1.49). The live101 SR 1.53 belongs to the raw-12m config, not the live residual-momentum one. No new ticket; D1 remains shadow-gated as documented.
2. Plain EWMA volatility targeting (no market state) improved eqmom walk-forward risk (Sharpe 0.76 → 0.95, maxDD −45% → −22%) but flat holdout Sharpe with a large return give-back — reconcile against the prior D4/BS-C result (book-matching target → constant SR) before use.