Quant Analyst — Feature-Ranking + Rules for a LuxAlgo Algo-Trading Dataset

PROJECT I run a live day-trading system on 120 US equities that consumes LuxAlgo Smart Money Concepts (SMC) alerts via webhook and converts them into trade decisions. I have 17 trading days of simulation data ready for analysis. I need a quant contractor to find features + rules that improve PnL. The system is currently ~breakeven across 17 days (-$32 net across 17,236 trades). My internal analysis has already identified some strong signals (e.g., "OB inside-zone × direction" is worth ~$7.5K in blocked bad trades), but I want a fresh independent read — no assumptions from my prior conclusions. DATASET ~17,236 rows (one row per trade) ~500 columns per row, joined from four tables: sim_transactions — trade identity + outcome (pnl_dollars_signal is the continuous target, was_win is the binary target) batch_decisions — 137 columns of state at entry decision time: SMC bias/scores, price action (ADX, RSI, VWAP position, ATR%, relative volume), momentum, price momentum, regime, sequence pattern matching, signal history, order-block state (inside_bull/bear, nearest OB coords, count above/below, pivot zone), BOS/CHoCH state (last direction/age/streak/trend regime) tdv_simulation_data — ~280 columns of per-symbol OHLC + indicators at entry (multi-timeframe: 1m, 5m, 30m, 60m, 1d; includes RSI, ADX, MACD, EMA, VWMA, BB, Ichimoku, pivots, ATR, sector performance) tdv_data_log — market-wide state at entry (SPX, DJI, VIX, breadth advance/decline, per-sector 5m performance) Also joined: exit-side context (do NOT use as entry features — look-ahead leak; useful for exit-timing analysis) Format: CSV export ~50-100 MB Data is delivered as one CSV. Column dictionary provided. WHAT I WANT Structured deliverable, roughly in this order of priority: Univariate PnL ranking of every feature. Cross-tab each feature (bucketed appropriately) against direction × PnL. Rank by absolute $ impact. I need a spreadsheet. Multi-feature rule mining. Find rule combinations of 2-4 features that would have vetoed losing trades (or sized up winning trades). Target: rules that beat any single univariate rule and have ≥50 trades supporting them. A single ML model (gradient boost / logistic regression, your choice) predicting was_win at entry time. Report feature importances. Baseline against a "trade all" baseline. I care more about interpretable feature-importance than raw AUC. A short written analysis — 5-10 pages — walking me through the top 10 findings, with charts and confidence intervals. Assume I'm technical but not a quant. WHAT I DONT WANT A trading strategy or backtesting framework I have my own A trained model I can't inspect A dashboard or UI Recommendations based on less than 30 trades of support SKILLS NEEDED Python + pandas + a boosting library (LightGBM / XGBoost) Comfort with tabular EDA at 17k × 500 scale Statistical rigor — confidence intervals, sample-size discipline, awareness of multiple-testing issues Bonus: prior experience with market microstructure or SMC/ICT frameworks (but not required — the data is self-describing) SCREENING QUESTIONS Please answer briefly in your proposal (skip generic templates — I read them): Send one example of prior quant EDA work (screenshot, gist, or Kaggle notebook is fine) In one sentence, how would you handle the fact that 33% of trades sit in the "trend regime = null" bucket (state hadn't been derived yet)? What's your process for guarding against look-ahead bias in tabular trading data? If a single feature slice shows +$50/trade edge on 40 trades, do you report it? Why or why not?