Trade NQ; use ES to identify the likely resolution mechanism. Date: 2026-08-09 (v2.4) · Data: NQ/ES front-month 1m bars 2021-01 → 2026-08 rebuilt from raw native contracts (A0); raw 1s NQ bars for destination ordering and execution; CME trading-day sessions. Status: exploratory. All data through 2026-08-08 are development data. Nothing here supports deployment; the prospective clock starts only after a frozen revision.
A divergence between NQ and ES is evidence that the usual relationship has been disturbed, not a directional trade signal. For every 10-minute observation with exact native interval endpoints (195,845 events; 162,669 with an exact consecutive 180-minute grid) this study reconstructs future paths, clusters them into resolution archetypes, freezes a point-in-time destination graph, estimates archetype probabilities from prior context, and designs causal entries whose targets and invalidations are structural. R:R is an output.
Headline findings:
- The event-time divergence does not resolve predictably. Among |e|≥1 events the residual contracts toward zero by 60m on 49.8% of cases vs 49.0% for matched non-divergence controls; the event-time residual sign aligns with the future 180m stretch on 50.8% (|e|≥2: 52.5%) — coin flips. The R1/R2 labels are therefore unconditional post-anchor path archetypes, not catch-up mechanisms.
- Archetypes recur and are fold-stable: rotation 72.6%, failed edge break 20.2%, common continuation 2.8%, NQ-stretch reversion 1.9%, persistent repricing 1.7%, ES-stretch reversion 0.8%; sub-family ARI R1 0.75, R2 0.62, R4 0.87, R5 0.63, R6 0.62 (R3 0.40 kept coarse); medoids frozen on the ≤2025 training fold.
- First touches are not validated waypoints. Actual reach rates are 58–62% by archetype, and conditional on a first touch the probability that the next structural target is reached before the adverse barrier is 0.23–0.63 (R5 0.63, R3 0.58, R1/R4 ≈0.25). Waypoints describe where paths go, not tradeable stepping stones.
- Context carries a small amount of archetype information, living in ES state and the developing path. Ladder (2026H1, purged): B0 0.7615 → B1 (prior-session structure + current-session state) 0.7577 → B2 (+ES raw state, relational vol/corr) 0.7542 ≈ B3 (+residual decomposition) 0.7529 → B4 (+developing path) 0.7342. Conventional Brier 0.4011 vs 0.4081 prior (skill 1.7%); top-1 equals the always-rotation baseline; the composite gate score reaches 8.6–9.8% precision at the top deciles vs a 4.3% base. B4 is response-side nowcasting, not a live-update justification. The object P(Z,L,τ|X) is modeled only for Z; destinations and times are descriptive incidence tables.
- No execution sleeve shows an edge under the proposal/entry/position state machine (entry-time arbitration, strictly-later retests, executable gap stops, timeout at the first print at/after the horizon, trade-level ledger): catch-up −0.252R/trade (CI −0.273,−0.229); leadership −0.067R (−0.093,−0.042), 2026H1 −0.108R (−0.175,−0.034); failure-fade −0.101R (−0.119,−0.086). Jul–Aug leadership is +0.055R on 28 trades — no evidential weight.
The framework is a market-description and hypothesis-generation tool. Every economic hypothesis tested is negative or unresolved; the prospective program (§10) is the only path to a strategy.
event or anomaly → complete future-path atlas → resolution/path clusters → natural destinations → pre-event context associations → archetype probabilities → causal entry trigger → structural stop and realized R:R.
The object estimated by a classifier is P(Z|X_t); destination and time layers are descriptive. Fixed-R excursions appear only as volatility units.
- Pipeline origin: A0 rebuilds native front-month 1m series, the additive continuous series and the 10m ledger from the raw cache; A1 computes relational returns on native contracts over the exact [t, t+10m) interval and rejects events lacking exact native endpoints; same-time-of-day statistics lag within slot before rolling.
- Sessions: CME trading days (17:00 America/Chicago roll; 16:00–17:00 CT break); RTH 08:30–15:00 CT; overnight 17:00–08:30 CT. Prior-session levels from the last completed trading day; internal-level touch histories require a true straddle; all "current" state uses bars whose interval elapsed before the decision.
- Paths: anchored at the known event close; timestamp-bounded windows with strict
< horizonmasks; max up/dn and their times are elapsed-minute quantities; fixed-u coordinates are NaN when no bar completed by the mark (missingness preserved through modeling); only rows with an exact consecutive 180-minute grid (162,669) enter the taxonomy; residual return-space denominators use native decision-time closes. - Destinations: internal (true midpoint/edges, session VWAP/extremes, prior close/VWAP, PROFILE_PROXY_POC/VALUE) and external (prior-session/overnight high-low, elapsed 90m swings); first-arrival and exit times in elapsed minutes; simultaneous edge hits rejected; candidates tied within a minute, and any destination sharing its minute with the edge exit, are resolved on 1s prints in consistent raw units with a strictly-later requirement, else discarded as ambiguous (A3 therefore also requires the raw 1s cache).
- Archetypes: defined solely from future-path behavior (stretch from the path's own first-60m residual maximum, closure attribution, efficiency, edge-break reversal); R1/R2 are post-anchor archetypes, not divergence-resolution mechanisms (§3).
- Learning protocol: expanding-year purged walk-forward; 6h-embargo variant; no scaler (HGB handles NaN); taxonomy and medoids frozen on the ≤2025 training fold and projected onto all rows; gate cutoffs per year from strictly prior years' out-of-fold score pools; ledger starts 2023.
- Execution authority: proposal/entry/position state machine on 1s data: proposals are generated at decision time, actual entry times are computed from the path, the one-slot position arbitrates by entry time; retests must occur on a strictly later second than the progress second; entries commit at the next observed open (no silent cancels; entry delay reported); stop gaps exit at the executable price; targets are resting limits paying actual G/R, never firing behind the fill; timeout at the OPEN of the first print at/after the horizon; a trade-level ledger (entry/exit timestamps, reason, target, stop, costs) is persisted (results/trades.csv).
- Metrics: net R and points per trade with day-block bootstrap CIs; hit rates from exit reasons; average winner = mean of per-trade G/R.
With the residual defined at the event in native points, |e|≥1 events (17,732) contract toward zero by 30/60/120/180m on 49.6–49.8% of cases; matched same-TOD controls with |e|<0.5 do so on 49.0–49.1%. The event-time residual sign predicts the future 180m stretch sign on 50.8% (|e|≥2: 52.5%). There is no divergence-resolution effect to exploit; the archetype taxonomy describes what paths do after an arbitrary anchor, and the divergence ledger remains a measurement anchor, not an entry.
Over exact 180m grids: an event edge breaks on ~87% (up) / ~85% (down); median time to the maximum up-excursion is ~96 minutes; median directional efficiency is 0.15; residual crossings at 5m sampling mark rotation as the default state.
Figure 4.1 — resolution archetypes and mean dimensionless NQ paths.
Mix over exact-grid rows: R6 72.6 / R4 20.2 / R3 2.8 / R1 1.9 / R5 1.7 / R2 0.8%. Sub-family stability (ARI between chronological fold fits, medoids frozen on ≤2025): R1 0.75, R2 0.62, R4 0.87, R5 0.63, R6 0.62; R3 (0.40) stays coarse.
Actual reach rates by archetype are 58–62% (unreached rows retained). Conditional on a first touch, the barrier-ordered probability that the next structural target in the travel direction is reached before the adverse barrier is: R5 0.63, R3 0.58, R2 0.47, R1 0.27, R4 0.23, R6 0.23; non-arrival shares within the horizon are 0.3–10.5%. Median MAE before the first touch is 3–5 points. First touches therefore describe path geometry; only R3/R5 show even moderate continuation-through-waypoint behavior, and none of it is traded profitably (§9).
Figure 6.1 — competing first-destination causes by archetype (180m, train folds).
Blocks (all past-only): B0 time/vol/NQ geometry; B1 prior-session structure plus current- session state; B2 + ES raw state and relational vol/corr; B3 + residual decomposition; B4
- developing path prefix (scientific only).
Figure 7.1 — context ladder (2026H1, purged).
Results: B0 0.7615, B1 0.7577, B2 0.7542, B3 0.7529, B4 0.7342; analogue (25-NN, smoothed) 0.9391; 6h embargo leaves B3 unchanged; expanding folds improve 0.912→0.740. Baselines: train-prior logloss 0.7832; always-rotation top-1 0.742 vs model 0.742; conventional Brier 0.4011 vs 0.4081 (skill 1.7%). Calibration reliability bins show rare archetypes under-predicted; common archetypes moderately calibrated. The composite S2-gate score scores 8.6–9.8% precision at the top 20–10% vs a 4.3% base — thin ranking, low recall. Reading the ladder honestly: structure (B1) adds little; ES state plus relational vol/corr adds −0.008; the residual decomposition adds nothing beyond that; the developing path adds −0.020 as response-side nowcasting and does not by itself justify live updates.
- S1 catch-up (R1/R2 gate): sweep–reclaim of the stretch-side extreme, fade to the event midpoint; stop beyond the swept extreme.
- S2 leadership (R5/R3 gate): breakout, progress beyond 0.3R, then a distinct held retest on a strictly later second (timestamp-verified five-minute closes respecting the edge; entry at hold completion's next open); stop at retest failure; target = nearest external destination ahead of the confirmation close; simultaneous edge hits rejected.
- S3 failure-fade (R4 gate): edge break that closes back inside the range; fade to the midpoint; stop beyond the broken extreme.
- R6: flat.
Gates: per-year 80th-percentile cutoffs from strictly prior years' out-of-fold scores. Proposals arbitrate the single slot by actual entry time.
Figure 9.1 — state-machine replay economics. (a) net R per trade by sleeve and period with bootstrap CIs; (b) structural target-hit rates.
| sleeve | period | trades | hit | avg win (R) | net R/trade | 95% CI |
|---|---|---|---|---|---|---|
| S1 catch-up | full | 17,188 | 38.3% | 1.57 | −0.252 | −0.273, −0.229 |
| S1 catch-up | 26H1 | 1,229 | 39.0% | 1.52 | −0.161 | −0.236, −0.099 |
| S2 leadership | full | 8,059 | 48.9% | 0.85 | −0.067 | −0.093, −0.042 |
| S2 leadership | 23–25 | 7,090 | 49.0% | 0.86 | −0.062 | −0.088, −0.032 |
| S2 leadership | 26H1 | 941 | 47.6% | 0.75 | −0.108 | −0.175, −0.034 |
| S2 leadership | 26JulAug | 28 | 60.7% | 0.58 | +0.055 | 0.023, 0.134 |
| S2 (B1 gate) | full | 7,918 | 48.4% | 0.85 | −0.077 | −0.103, −0.050 |
| S3 failure-fade | full | 12,851 | 53.7% | 0.87 | −0.101 | −0.119, −0.086 |
| S3 failure-fade | 26H1 | 1,682 | 53.2% | 0.88 | −0.061 | −0.107, −0.020 |
Every full-period and 2023–25 sleeve is negative with CIs excluding zero; 2026H1 leadership is negative as well; the Jul–Aug figure rests on 28 trades and carries no evidential weight. Average winners pay 0.75–0.9R against 1.0R losers plus 1pt costs. The B1-gated variant is indistinguishable from the B3-gated one. Entry delays are median 0s (max ~1h across maintenance gaps, reported per trade).
No deployable edge is claimed. The 30–40%-win / 3–5R-winner profile appears nowhere.
- All numbers are development; gates and rules were searched on this cache; family-wise multiplicity adjustment (PBO / Harvey–Liu–Zhu) is owed before any promotion.
- Touch⇒fill on resting targets remains optimistic; FIFO queue position, latency and impact are unmodeled; tick/MBO replay is the first required robustness test.
- The prospective clock starts after a frozen revision, on data generated after the freeze; OOF folds cannot undo repeated research on the same 2021–26 cache.
- Ranked follow-ups: (1) freeze + prospective confirmation with pre-set kill thresholds; (2) tick/MBO replay with queue-aware fills and bid/ask-side entries; (3) path-conditional updates (B4) to time entries; (4) multiplicity and block uncertainty; (5) portfolio ledger (dollars, turnover, exposure, drawdown); (6) true volume-profile destinations; (7) a divergence-conditioned strategy would need a resolution effect, which §3 shows is absent — future work should treat ES–NQ state as volatility/nowcasting context only.
- Dynamic residual: z(r^NQ) − β_slot·z(r^ES), β from prior 25 sessions per slot.
- Resolution archetypes R1–R6: future-path descriptions; R1/R2 are post-anchor archetypes, not resolution mechanisms.
- PROFILE_PROXY_*: OHLCV-volume approximations of profile levels, named as such.
- Purge/embargo: train rows whose 180m label window reaches (purge) or comes within 6h of (embargo) the validation start are excluded; windows timestamp-bounded; only exact consecutive grids enter the taxonomy.
- State machine: proposals at decision time; slot arbitration by actual entry time; committed next-open market entry; executable gap stops; resting-limit targets paying actual G/R, never firing behind the fill; timeout at the first print at/after horizon; trade-level ledger persisted.
- R / G: R = |entry − structural stop|; G = |target − entry|; R:R = G/R, an output.
Self-contained bundle: scripts A0–A9 + v2_common; A0 rebuilds natives, continuous series and the 10m ledger from the raw cache; shipped derived inputs (data__) and results (results__, including the model pickle and trades.csv), placed by setup.sh; DATA_MANIFEST.md hashes every consumed raw file (1,541 entries) plus derived artifacts; verify_manifest.py --cache checks every record by relative path; DATA_SOURCES.md documents retrieval and scope (A3 and A7 require the raw 1s cache, hash-pinned but not shipped); requirements-lock.txt carries macOS and manylinux wheel hashes; ENVIRONMENT.txt; RUNBOOK.md lists setup.sh as step 0 and the full chain; SHA256SUMS covers scripts, docs, charts, manifest and shipped artifacts; V2_CACHE env override; scripts create their directories.
Protocol: every statistic is computed from information available at the decision time unless labeled a response-side description; future paths define historical labels only; no horizon, cluster count, gate or exit was selected from payoff; nulls are reported as nulls.
