How a signal qualifies

Each evaluation requests up to 450 finalized OHLCV bars on each of three horizons, plus current funding, current and prior open interest, and the nearest 200 bid and ask levels from the public OKX production API. V4 resolves liquidity separately at 5, 25 and 200 levels, calculates top-of-book microprice pressure and records exchange-data age. Every candle request is aligned to its last completed interval so forming candles are excluded. There is no synthetic or demo fallback: unavailable, incomplete, stale or malformed live data fails closed.

• Default scope: live linear USDT perpetuals in OKX crypto, stocks/ETFs, commodities, forex and bonds; categories with no eligible listings remain empty
• A category-interleaved rotation covers the configured catalogue in bounded batches; catalogue inclusion does not imply signal qualification
• Official OKX instrument categories—not ticker-name guesses—define asset class
• Default stack: 1H trigger, 4H setup and UTC-aligned 1D regime
• Up to 450 finalized bars per timeframe; minimum 120 required
• 200 order-book levels per side
• 5/25/200-level imbalance, microprice and spread telemetry
• Maximum order-book age: 30 seconds
• No exchange trading keys or account access
• One fingerprint per symbol, bar close and model version

The engine evaluates long and short through the same function. A causal four-state diagonal-Gaussian hidden Markov model estimates bear, neutral, bull and high-volatility state probabilities from return, true-range and volume observations on each horizon. Only filtered probabilities available at the decision bar are used. V4.1 also measures posterior persistence and transition risk. The daily regime and four-hour structure must align, while the one-hour trigger may be supportive or a controlled pullback but cannot materially oppose the swing thesis. A long now requires a daily HMM bull probability of at least 60%, while a short requires a daily HMM bear probability of at least 40%. These are qualification thresholds, not forecasts of trade success.

• EMA20 versus EMA80 trend on all three horizons
• Daily trend persistence across the latest 30 bars
• Causal HMM state probabilities fitted independently on each horizon
• Inclusive daily HMM gate: bull ≥60% for longs; bear ≥40% for shorts
• Kaufman-style signed trend efficiency and 50-bar log-trend R²
• Short-versus-long realized-volatility and true-range shock vetoes
• Four-hour three-, five- and ten-bar momentum blend
• Four-hour close against a 50-bar channel calculated through t−1
• One-hour trigger opposition and pullback controls
• Open-interest change used only as participation confirmation
• Funding crowding is amplified when open interest expands
• Conflicting 5/25/200-level liquidity can veto a setup

The quality score is a weighted evidence composite between 0 and 100. It is not presented as a win probability. V4 assigns 12% to deterministic horizon agreement, 11% to HMM horizon agreement, 7.5% to the daily trend regime, 9% to the daily HMM state, 8% to setup trend, 7% to trend efficiency, 4% to trend linearity, 4.5% to regime persistence, 6.5% each to momentum and breakout structure, 3.5% to trigger quality, 4% each to directional volume and multi-depth liquidity, 2% to microprice pressure, 2.5% to open-interest participation, 2% to swing location, 2.5% to volatility quality, 1.5% to funding crowding and 2% to regime-transition stability.

• Minimum publish score: 66/100
• Maximum default spread: 8 basis points
• No heuristic expected-value claim
• Every contribution is stored in a score ledger
• Every alert exposes its leading drivers and weakest qualified factors
• HMM stress, volatility-ratio and range-expansion limits fail closed

Entry is a reference based on finalized market data, not a guaranteed fill. The invalidation distance is the greatest of 1.8 four-hour ATR, 0.75% of price or a capped recent swing-structure distance with an ATR buffer. Targets are placed at 1.5R, 2.5R and 4R. Entry validity expires after three four-hour bars while outcome evaluation runs for 72 hours.

• Subscribers must account for spread, slippage, fees and funding
• Actual fills will differ from the reference
• Position size is deliberately not prescribed
• The optional calculator uses only subscriber-entered assumptions and stays client-side
• No averaging down or liquidation-distance assumptions

Signal IDs, model versions, levels, feature snapshot, score ledger and generation times are immutable. Email delivery is leased and idempotent per signal and entitled recipient. Telegram is at-least-once and every V4 alert carries its immutable signal ID, model, source, finalized-bar time and live decision matrix. Outcome scoring uses the originally published levels and accounts for publication delay.

• No backfilling signals after price moves
• No quietly moving stops or targets
• No probability label before chronological prequential calibration and manual promotion
• Model changes require a new version and shadow test

The selection overlay retains 73 finalized setup closes per successfully acquired instrument across scans. It computes synchronized log-return correlations over 24 and 72 bars plus an exponentially weighted 72-bar estimate with a 12-bar half-life. On the default 4H setup these represent four days, twelve days and a two-day half-life. Estimates are rolling descriptive statistics, not trained return forecasts. Timestamp gaps, constant series, insufficient history, mismatched timeframes and stale inputs remain unavailable.

The reference book contains unresolved OKX signals queued for dispatch or delivered within the last 72 hours, not confirmed customer positions. It conservatively reserves queued exposures and removes signals once outcomes resolve or the research horizon elapses. A 64-exposure bound fails closed rather than silently truncating the portfolio. A repeat symbol is held, including an opposite-side candidate; it is not assumed to close an existing trade.

Candidates that already passed the baseline quality gates are greedily ranked by lowest maximum directional correlation to the book, then by quality and stable symbol order. After each selection, the next candidate is re-evaluated against the enlarged book. A signed correlation of at least 0.80 on any of the three estimates prevents an addition. Missing book comparisons also hold publication. Selected candidates retain their original quality score and trade levels. Decisions, reference peer correlations and policy version are recorded before persistence; held candidates are not published or fed into the existing outcome learner as executed trades.

The broader cached-market comparison counts low absolute overlap only when all three absolute correlations are below 0.30. It is context, not an allocation optimizer or a claim of uncorrelated strategy returns. Thresholds are initial conservative heuristics and require prospective validation. There is no covariance-matrix inversion, inferred expected return, leverage recommendation, position sizing or trade execution. The outcome learner’s numerical feature vector remains unchanged; portfolio diagnostics are stored separately.

• Rolling histories refresh from the same genuine OKX data used by the engine
• No synthetic filling and no missing-data diversification credit
• Queued and unresolved signals are a proxy, not subscriber holdings
• Cross-asset category labels alone do not establish diversification
• Correlation can rise or change sign during stress; low historical overlap is not guaranteed future independence

A separate rolling evidence layer groups only unambiguous live OKX target-or-stop outcomes whose signals were generated during the previous 72 hours. Every cell is exact to instrument, side and a coarse decision-time scenario: side-relative HMM alignment, an information state from directional volume plus 5/25/200-level book imbalance and microprice pressure, and compressed, normal or expanded volatility. Scenario labels come from the immutable feature snapshot, so later market data cannot rewrite the state that was known when the signal qualified.

Raw win rate is reported for audit, but ranking never uses it directly. Each exact cell is shrunk toward a leave-cell-out baseline for the same instrument and side, must contain at least eight resolved outcomes, and must place an 80% posterior evidence band entirely above or below 50% before it can influence order. The adjustment is capped at 2.5 quality points and applies only after deterministic qualification, as a bounded portfolio-ranking preference. It cannot create a candidate, reverse direction, bypass a veto or change entry, stop or targets. Missing, malformed, ambiguous and timed-out evidence is neutral.

Coarse bins and shrinkage reduce variance and scenario memorization; the exact-pair requirement prevents one market from lending an apparent edge to another. The minimum sample and neutral fallback protect against overfit, while the same-pair partial prior and deliberately coarse states reduce underfit from excessive fragmentation. This is short-horizon adaptive evidence, not a profitability estimate, and its policy version is stored with each portfolio decision.

• Three-day window is recomputed causally at every scan
• Long and short evidence remains separate for every instrument
• Only binary, unambiguous target/stop labels enter the report
• Raw and Bayesian-shrunk rates plus the evidence band remain visible to the operator
• Maximum preference: ±2.5 quality points; original published quality remains immutable

The qualification engine remains reproducible: HMM fitting uses fixed initialization, a fixed iteration budget and only the finalized history available at the decision time. A separate six-head learner combines one global FTRL-Proximal logistic model with five deterministic online-bagged FTRL members. Every member prediction is frozen before the outcome exists, then the ensemble updates only from finalized live-market outcomes under a versioned label contract. It starts in shadow mode and cannot change eligibility, direction, entry, stop or targets.

Outcome evaluation starts after a configured publication delay, uses finalized OKX lower-timeframe bars with a next-open execution assumption and explicit slippage, censors same-bar target/stop ambiguity, and excludes non-OKX, timeout and ambiguous observations from production training. Probabilities use causal Bayesian bin calibration and are benchmarked against the pre-outcome empirical success rate—not the public quality score. Promotion is manual and requires at least 500 resolved outcomes, 100 outcomes per class, 200 recent forecasts, Brier and log-loss improvement, calibration error at or below 8%, AUC of at least 0.55, bounded ensemble disagreement and no Page-Hinkley or fast/slow loss drift. New model and feature-context IDs prevent legacy observations from entering this learner. Any active percentage is conditional on an unambiguous target-or-stop resolution inside 72 hours; timeouts are excluded and it is not a profit forecast.

• No learned-score display before promotion readiness
• No black-box override of hard gates
• Five online-bagged members expose model disagreement
• Dual concept-drift monitors automatically suppress active scoring
• The core decision remains reproducible from stored inputs