How a trade actually
gets placed.

Survivors are evaluated across three timeframes. The 4h trend is a hard gatekeeper — if it disagrees with the trade direction, the symbol is eliminated regardless of any other signal. 1h confirms. 15m times the entry. Cached results mean this runs at zero API cost on warm cycles.

Three tiers built from live win-rate data across 32+ pre-v2 trades. Tier A (AVAX, WIF, SUI, CRV, SNX, NEAR, MAG7, SEMIS, BNB, LINK, INJ, JTO, APT, ROBOT, INFOTECH) — scanned every iteration; proven performers with 100% WR in live data. Tier B (SOL, XRP, DOGE, NVDA, TSLA, META, COPPER, PLATINUM…) — scanned every 3rd iteration; neutral or mixed results. Tier C (BTC, ETH, AAPL, AMD, GOLD, BRENTOIL, CL, NATGAS, COIN, MSTR, AAVE, SEI) — permanently excluded; 0% win rate in live data. Data-driven exclusion: our indicator suite produces no edge on these assets. Re-evaluated automatically every 50 trades.

Before triage, every symbol is enriched by four independent passes running concurrently. VWAP filter: rolling 24h VWAP — signals opposing VWAP are flagged CONFLICT. Pre-market range (stocks only): pre-market high/low (4:00–9:30 AM ET) + first-30-min range — breakout direction matched against signal. Volume delta: buy vs sell pressure from the last 20 candles, including BULL_DIV and BEAR_DIV reversal flags. Momentum indicators: MACD (12/26/9), Stochastic RSI, Bollinger Bands + Keltner Channels, and TTM Squeeze. When BB bands contract inside Keltner Channels the symbol is coiling; when they expand back out the squeeze fires. Market structure: 4h swing high/low analysis labels every symbol HH/HL (uptrend intact), LH/LL (downtrend intact), or BROKEN (structure just broke). Liquidity sweeps: detects stop-hunt wicks on 1h candles — when price wicks below a swing low then closes above it (SWEEP_LOW), institutional accumulation has just occurred. A sweep aligned with the MTF signal bypasses all other filters and is guaranteed to pass triage. An opposing sweep is a hard reject. OBV + CVD: On-Balance Volume and Cumulative Volume Delta track institutional flow — when both are rising while price falls (dual bull divergence) or falling while price rises (dual bear divergence), the signal is STRONG_BULL or STRONG_BEAR and ranks just below a confirmed sweep. Supertrend + Pivot Points: ATR-based trend direction line (period=10, multiplier=3) that flips on trend reversal (BULL_FLIP/BEAR_FLIP = highest-conviction fresh signal); Classic pivot points from the prior 24h session provide S1/S2/R1/R2 price levels — a Supertrend flip + pivot break is a BREAKOUT signal. All eight layers are fail-open, zero extra API calls for OBV/CVD, Supertrend, or Pivots.

All 92 symbols scored in a single batch call to DeepSeek. Prunes low-probability setups before any expensive model runs. Symbols that don’t cross the minimum confidence threshold are dropped here — instantly, at near-zero cost.

Open interest and funding rate checked for every candidate. Extreme crowded-long or crowded-short scores (> 0.7 or < −0.7) flag the setup as a crowded trade risk. Position sizing is reduced or the setup is skipped.

Hyperliquid liquidation WebSocket monitored continuously. When forced-selling exhaustion is detected — a $1M+ cascade that suddenly stops — the detector fires a bottom-signal for that symbol. This is a separate alpha source that can open trades independent of the main pipeline.

18 directional keyword buckets (~300 keywords) pre-screen headlines from 6 sources every 90 seconds. Buckets are explicitly signed — bearish_oil catches G7/IEA reserve releases, ceasefire, Iran deal; bullish_oil catches OPEC cuts, Hormuz closure, pipeline attacks. DeepSeek classifier applies 30+ explicit event→asset-bias rules and returns per-asset directional biases. Macro context is injected directly into DeepSeek batch strategy prompts, filtering setups before analysis rather than only after.

Three agents independently evaluate each candidate. The Strategy Agent (DeepSeek, 50% weight) receives full enrichment per symbol — MACD, OBV, CVD, Supertrend, Market Structure, Liquidity Sweep and news context, all wired in for the first time in v2. The Technical Agent (40%) evaluates indicator confluence locally. GPT-4 Vision (10%) renders a live chart and now receives regime context + strategy action for richer pattern analysis. After combining scores, news/regime confidence penalties are applied: −8% for UNKNOWN regime, −10% for major news events, −5% for risk-off macro. Platt scaling calibration then corrects model overconfidence against live outcomes — activates automatically at 50 closed v2 trades.

All finalists sent to Claude in a single API call. v2 change: Claude must now write the strongest possible bear case against each trade before issuing a verdict. The default disposition changed from “approve unless reason to reject” to “scrutinise, then decide.” Per-candidate enrichment block, conflict flags, drawdown %, and confidence scores are all visible to Claude. Three verdicts: APPROVE (trade fires at full size), REJECT (trade dropped, bear_case + reason logged), or REDUCE_SIZE (0.25–0.75× multiplier). Rules: 3+ indicator conflicts = REJECT unless R:R > 2.5:1; drawdown > 8% = REDUCE_SIZE on any marginal setup; >3 same-direction positions = automatic REDUCE_SIZE. Fail-open on timeout. All verdicts, bear cases, and reasons stored in scan_log.json and visible in the dashboard Signals tab.

Five independent guards run in sequence after Claude approval. Any block terminates the entry. Zero-score guard: blocks entry if both strategy_confidence and technical_confidence are 0.0 (scoring pipeline failed silently). Unknown regime guard (v2): blocks all new entries when crypto regime is UNKNOWN — data showed UNKNOWN regime produced 31% WR and −9.54 in 16 pre-v2 trades. Drawdown confidence boost (v2): requires ≥75% confidence when drawdown exceeds 8%. Regime alignment guard: per-asset-class — STRONG_BULL blocks shorts in that class, STRONG_BEAR blocks longs. Does not cross asset classes (crypto bull cannot block oil shorts). R:R gate (v2): TP distance / SL distance must be ≥1.5. Computed from actual entry price after stops are finalised. Data-driven: pre-v2 R:R was 0.53× (avg win .25 vs avg loss .14) — mathematical guarantee of loss at any win rate below 65%.

The stacked ensemble (GradientBoosting + RandomForest + Logistic Regression) produces a final win-probability score using historical features: symbol-specific win rate, market regime, agent confidence, funding environment, volatility regime, and more. SHAP values explain the score. The meta-learner can veto a trade even if all three agents agree.

Position size is computed from a base Kelly amount multiplied by 7 independent factors, each drawn from the enrichment_snapshot saved on the opportunity during the consensus stage. Volatility (ATR%): ATR <1.5% of price → 1.1×; ATR >5% → 0.6×. Regime: STRONG_BULL/BEAR → 1.15×; UNKNOWN → 0.65×. Category: commodity → 0.70× (macro-driven, less predictable); stock → 0.80×; crypto/pre-IPO → 1.0×. OI liquidity: open interest >0M → 1.05×; <M → 0.70×. Indicator alignment: all 3 (VC + Supertrend + MACD) agree → 1.10×; 0/3 → 0.80×. Combined multiplier is floored at 0.20× and capped at 1.20× to prevent extremes. Dynamic TP widening: when 3 indicators align strongly, take-profit target is widened to 7× ATR (vs default 4×) — letting winners run when the setup is genuinely strong. Data: pre-v2 take_profit exits averaged +5.70 (100% WR) vs stale exits +.90 — the problem was TPs set too tight.

All v2 guards have passed. The order fires on Hyperliquid or Binance via the exchange router. ATR stop and take-profit are set simultaneously at entry (4–7× ATR depending on indicator alignment). Trailing stop activates after a 2.5% favourable move. Position tracked in PostgreSQL. Trade grade (A–D) computed on close and fed to the reflection agent and meta-learner. Symbol Protection V2 updates win/loss streaks per symbol — 3 consecutive losses triggers a 3-day block.