Qc Deep Feature Forensics

12-dimensional technical feature attribution engine — compares winner vs loser trade entry conditions using RSI, Bollinger, MACD, volume surge, gap, and more to find what makes winning entries different.

Go beyond P&L analysis. This skill answers: "What market microstructure conditions were present when winning trades were entered vs losing trades?"

When to use

• "Why do some of my trades win and others lose?"
• "What entry conditions lead to profitable trades?"
• "Run a feature attribution on my backtest"
• "What-if analysis: would filtering out RSI > 70 trades help?"
• After running qc-order-forensics for high-level diagnosis, use this for deep-dive

How it works

python3 deep_forensics.py <orders.csv>

Pipeline

1. Order Reconstruction: Groups raw orders into closed trade pairs (buy group + sell group per contract)
2. Batch Data Download: Fetches historical daily OHLCV from Yahoo Finance per ticker (cached to local CSV to avoid re-downloading)
3. Technical Indicator Pre-computation: Calculates all 12 indicators on the full ticker history
4. Feature Extraction: For each trade, extracts a 12-dimensional feature vector at the entry date
5. Winner vs Loser Comparison: Statistical dual-sample comparison with diagnostic interpretation
6. What-If Simulation: Tests hypothetical filters and measures their net impact on total P&L

12 Feature Factors

Factor	Description	Format
gap_pct	Gap open % vs previous close	%
volume_surge	Volume / 10-day average volume	x
ma5_deviation	(Close - MA5) / MA5	%
ma20_deviation	(Close - MA20) / MA20	%
volatility_expansion	Intraday range / 5-day ATR	x
intraday_return	(Close - Open) / Open	%
rsi_14	RSI(14)	0-100
bb_position	Position in Bollinger Band (0=lower, 1=upper)	0-1
macd_hist_norm	MACD histogram / Close	ratio
consecutive_up_days	Count of consecutive up-close days	days
distance_from_20d_high	(Close - 20d High) / 20d High	%
prev_day_return	Previous day's return	%

Report Output

Section 1: Feature Mean Comparison (Winners vs Losers)

Table showing each factor's mean for winning vs losing trades, the delta, and a diagnostic interpretation. Example:

| Gap Open %    | +3.2% | +1.1% | +2.1% | Winners enter on stronger gaps |
| Volume Surge  | 2.8x  | 1.4x  | +1.4x | Volume confirmation helps      |

Section 2: What-If Filter Analysis

Simulates applying each filter rule to the full trade set and measures:

• How many losers would be avoided
• How many winners would be accidentally killed
• Net impact on total portfolio ROI
• Verdict: Shield (improves total P&L), Toxic (kills outlier wins), or Marginal

Filters tested include: gap > 2%, volume > 2x, below MA20, RSI > 70, BB > 0.95, consecutive up > 3, near 20d high, negative intraday return.

Combined filter: Stacks all "Shield" filters and reports the combined effect.

Section 3: Winner Entry Profile

Statistical percentile ranges (25th-75th) for each factor among winning trades. Defines the "ideal entry environment" envelope.

Output Files

• <name>_features.csv — Full feature matrix for all trades
• feature_diagnosis.md — Complete markdown report

Caching

YFinance data is cached per-ticker in yfinance_cache/ alongside the orders CSV. Subsequent runs on the same data skip downloads entirely.

Dependencies

pip3 install pandas numpy yfinance

Rules

• Do not modify files in yfinance_cache/ manually. The cache uses date-range coverage checks. Corrupted cache files will cause silent data gaps in indicator calculations.
• Do not change indicator parameters (RSI period, Bollinger window, etc.) without understanding that all 12 factors are calibrated together. Changing one shifts the entire winner/loser comparison baseline.
• What-If verdicts ("Shield" vs "Toxic") are based on total portfolio ROI impact, not win rate. A filter that improves win rate but kills outlier wins is marked "Toxic" because total P&L decreases. Do not override this logic.
• Internet access is required on first run for each ticker. Subsequent runs use cached data. If you run in an offline environment, pre-populate the cache directory.
• Minimum 20 closed trades required for statistically meaningful feature comparison. With fewer trades, the report will still generate but conclusions are unreliable.