Warrior_EA/Meta_Labeling_Design.md

Meta-Labeling: the NN predicts trade quality, not market direction

Status: DESIGN — agreed 2026-08-12, implementation staged below.

Why this target

Per-bar direction of one instrument from its own H1 history is measured dead three detector-independent ways (MI noise floor, lag profile 0–20, post-backprop-fix retrain at chance vs a positive control that finds volatility at 4x its null). No topology change can extract information that is not in the input-target pair. Every credible "NN works on markets" result changes the target or the data instead. This design changes the target; cross-sectional pooled training (change the data) is the follow-up lever, and a quantile head can ride along on the retrain either forces.

The new question the NN answers: given that a specific setup just fired, will THIS instance reach its target before its stop, at the EA's own geometry, net of cost? That is a well-posed supervised problem where per-bar direction never was:

• the label is the realised outcome of an actual trade shape, not a synthetic class;
• the base rate is the setup's win rate (~40–65%), not a 1% tail;
• part of the answer is PROVEN predictable (range at 4x null; spread/vol/session state), because "does the target get hit first" is partly a volatility-geometry question;
• every labeled example is a moment someone would actually consider trading.

Architecture

classic patterns (26 x Buy/Sell)  ->  candidate: (side, pattern, geometry)      [WHEN]
                                          |
excursion head (proven)           ->  SL/TP geometry for the candidate          [SHAPE]
                                          |
META-HEAD (this design)           ->  P(win | candidate, market state)          [WHETHER]
                                          |
vote = f(P - break-even)          ->  existing threshold / MM / risk plumbing   [HOW MUCH]

• Primaries: the existing classic signals. They are NOT assumed to carry edge — they define when a candidate exists. SQX remains the user's separate rule factory; nothing here parses or reverse-engineers SQX. More primaries (e.g. the user's Wyckoff blocks) are added only after the meta-head demonstrates skill on the current 52.
• One net for all patterns, pattern identity as input features — shares representation across patterns, and event-rare patterns borrow strength from state-common ones.
• Integration point: the per-side pattern slots added in 652bf81 (GetActivePatternLong/Short + LastNetVote) are exactly the hook: the meta-signal reads which candidates fired this bar and emits a gating vote for that side. The AI filter's global direction vote (closed verdict) is what it replaces.
• Unchanged: excursion head, expectancy stop, risk budget, the DB as objective logger with its decision layer (the DB keeps ranking primaries; the meta-head is a second, richer consumer of the same "did this instance pay" question).

Training

• Sample: bars where at least one pattern fired, found by sweeping the classic ladders over training history (the ladders are already index-based; a historical sweep is cheap). One training row per (bar, fired pattern, side).
• Label: triple-barrier outcome for the proposed side at the EA's own SL/TP geometry — the existing labeler (b4a704d line), evaluated from the fire bar. Win=1 / loss-or- timeout=0. Same horizon rules as current training.
• Features: the existing BuildFeatureWindow() output (unchanged front end), plus a setup descriptor appended at the head: pattern one-hot (26), side, the filter's own netVote, SL and TP distances in ATR, spread/ATR at fire time.
• Head: binary logistic with the existing logit prior-correction (f2ec1ed), trained with the existing optimizer/batch-norm/mini-batch stack — all of that machinery is target-agnostic.
• Checkpoint selection: the existing operating-point objective, coverage x (precision − break-even) on the purged calibration slice (2189316, 983a6a3) — it was built for exactly this shape of question.
• Deploy gate: family-wise, as always — the final deploy decision is best-of-N over eras and must clear the null of the maximum (1df3054 lesson). No exceptions because the idea is likeable.

The honest risk, stated up front

If the primaries carry zero conditional structure, the meta-head degenerates into a volatility/cost/session timer. That is the FLOOR, not failure — "skip candidates whose geometry can't pay in this vol/spread state" is real risk control — but it is bounded and must not be sold as entry edge. The experiment's upside case is that sparse conditional structure exists at setup moments that the dense per-bar tests could not see (they measure marginal per-bar information; a signal alive on 3% of bars is invisible to them by construction). The gate decides which case we are in; nothing else does.

Staged implementation (each stage compiles and is testable alone)

1. S1 — corpus from the signal DB (implemented; supersedes the original "training-time ladder sweep"): the per-side journaling built in 652bf81/195be20 means the DB already IS the candidate stream — every instance the live ladders fire, both sides, uncensored, with netVote and entry price. Reading it (Expert\AIBase\MetaCorpus.mqh) instead of re-implementing 26 ladder conditions in training code eliminates the silent-divergence trap outright: the corpus is by construction identical to live behaviour. Costs accepted: corpus coverage = whatever backtest populated the DB (a corpus build is a tester run with UseDatabaseRanking=true and the new DB_MaxRowsPerTable input raised, e.g. 20000, so a 15–20 year run isn't pruned), and sampling is one candidate per fire-stretch (the correct dedup for training anyway). The VerboseMode OnInit report prints corpus volume per family, closed fraction, and the measured GMT→server bar-offset match table that S2's label plumbing pins to.
2. S2 — meta-head + label plumbing (implemented 2026-08-13): Signals\SignalMETA.mqh (AIType = AI_META) + meta seams through the existing training machinery. What shipped, including the four deviations from the sketch above and why:
   • 2-output softmax head, not a separate logistic path: a 2-class softmax+CE is mathematically a logistic/BCE head, so the total==3 host-side gradient in AI\Impl\NetForward.mqh was generalized to 2||3 (both backprop paths; zero backend changes) instead of growing a parallel loss.
   • Descriptor at the INPUT, not "at the head": CNet has no concat layer. On the S2 MLP front end input-append is equivalent up to one linear layer. 31 features: 26-slot pattern one-hot, side, tanh(netVote/20), SL/TP in ATR, spread/ATR at fire. Conv/LSTM meta variants would need whole-pseudo-bar padding — out of S2 scope.
   • No prior correction: the logit adjustment is 3-class-shaped and exists for a ~6% tail; the meta base rate is ~40% and the operating-point fit carries calibration.
   • Trained on the live chart like every other model (not "offline in the tester" — the tester is inference-only by architecture). The corpus is read at attach from the LARGEST signal DB on disk (config-fingerprint-independent — the S1 procedural traps came from exactly that coupling), resolved to server bars per era with the GMT offset MEASURED PER ROW against entryPrice vs bar open (DST-immune; histogram logged), and pre-2017 daily-backfill rows are dropped by a window-span regime filter.
   • The whole selection/deploy stack runs unchanged via a counter mapping (win→Buy, loss→Sell): precision = wins among traded candidates, chance = base win rate, buy/sell recall = sensitivity/specificity (kills always/never-call collapses), operating point = coverage x (win − geometric break-even) on the purged calibration band, deploy gate = family-wise as always. Era-end META log line reports cov x (p − BE) vs the base-rate null with its standard error.
   • Labels are the side-conditional triple-barrier win caches (never the DB's S&R outcome); the fingerprint gains a conditional |TGT:META1 token and meta models live under State\META\ with the 2-output filename slot, so no collision with direction models. S2 casts no votes: live inference/online-learning entries are guarded off for the meta target until S3.
3. S3 — inference integration: meta-signal consumes fired candidates via the per-side hooks and votes; classic weights stay as a fallback layer; VerboseMode panel exposure.
4. S4 — retrain + family-wise gate + (only if passed) deploy.

Cross-sectional pooling (one model over SP500 + the 4 tick + 5 SQX-decoded instruments, scale-free features already in place) is the next lever after S4, whatever S4's verdict. Snapshot/seed ensembling rides along from S2 (average the last K checkpoint snapshots instead of trusting one).

S2 verdict on SP500 H1, and the pre-registered H4 experiment (2026-08-13)

Measured over 350+370 eras: the head carries real, persistent ranking skill — the first out-of-sample signal since the pivot — but 0 eras cleared cov x (p − BE). The decomposition line localized it: the lift is LONG-ONLY (+1.8–2pp; the model's short picks do worse than the short base rate), strongest on the MA family (67–70% traded win, nominally above the 67.5% break-even but <1σ on ~350 trades and best-of-32 cells — not family-wise evidence). The clinching arithmetic: edge x width = 0.02 x 4.74 ATR ≈ 0.095 ATR/trade against a measured spread of 0.099 ATR/trade. The signal is real and being consumed exactly by cost.

H4 experiment (pre-registered before any H4 data exists): on SP500 H4 the ATR roughly doubles while the spread is unchanged, so the cost drag (BE − base gap) roughly halves (~1.3pp) while the lift — if it transfers — stays ~+2pp. Same pipeline end to end: S1 corpus backtest on H4, AI_META chart on H4. Hypotheses, stated in advance:

• H1 (primary): cov x (p − BE) turns positive at ≥25% coverage, clearing the 2σ edge floor — deployability decided by the existing gate, nothing relaxed.
• H2 (secondary): the lift remains long-only; SHORT traded win stays at/below its base.
• H3 (risk): the lift decays with timeframe as fast as the cost does (the tick-flow failure shape), leaving the net at zero again. If H3 is what the data says, the single-instrument well is closed at every accessible cost point and cross-sectional pooling is the only remaining lever. The known thinning: ~1/4 the bars, so ~11k usable candidates — acceptable for this head. The corpus loader now REQUIRES a symbol+timeframe-matching DB, so the H1 corpus cannot leak into the H4 run.

H4 RESULT (2026-08-13, 999 eras, 13,436 candidates): H3. The cost gap halved exactly as computed (BE 64.1% vs base 63.0% = 1.1pp at the derived 3.06/1.71 geometry) — and the lift did not come with it: max cov x (p − BE) = +0.07 (1 era of 999, pre-2σ), skill at selective thresholds +0.47pp mean (noise), and the head's ranking on H4 was slightly INVERTED out of sample (its skipped candidates won 67% vs 62% for its trades), so the expectancy fitter correctly pinned the threshold at 0/100% coverage. The lift shrank faster than the cost — the tick-flow failure shape, now measured at the setup-conditional level too. The single-instrument SP500 well is closed at both accessible cost points, with pre-registered hypotheses. Remaining levers, in expected-value order: (1) candidate streams that carry positive base edge (the user's own SQX strategies' signals as a journaled candidate family — the meta head is an amplifier, and it finally would have something to amplify; their trade-list CSVs measure this before any build), (2) cross-sectional pooling (the last pure-NN lever; honest prior lowered by the twice- measured signal≈cost equilibrium), (3) S3 as a variance-reducing risk filter at zero net edge.