🚀 EdgeRank

A machine learning system for identifying and executing high-confidence trading opportunities.

---

🧠 Overview

EdgeRank is a machine learning-driven trading system designed to identify high-probability short-term opportunities in equities.

Instead of predicting everything, EdgeRank focuses on:

Acting only when confidence is high and the best opportunity is clear.

The system has evolved into a hybrid architecture combining:

• 🎯 Per-ticker specialization
• 🏆 Cross-asset competition
• ⚖️ Strict portfolio selection

---

🎯 Core Philosophy

Train broadly. Act selectively.

• Models learn from historical data per asset
• Signals are filtered by confidence thresholds
• Opportunities compete across the market
• Only the best trade(s) per day are executed

---

🏗️ System Architecture

Data → Features → Models → Signals → Ranking → Trades → Backtest

1. 📥 Data Ingestion

• OHLCV market data (free sources)
• Multi-ticker dataset
• Cleaned + standardized

---

2. 🧮 Feature Engineering

Returns

• 1d, 5d, 20d

Trend

• SMA (5, 20, 50)
• Distance from SMA

Volatility

• Rolling std (5d, 20d)

Volume

• Volume averages
• Volume ratios

Targets

• Binary: next-day up/down
• Continuous: next-day return

---

🤖 Modeling

Baseline

• Logistic Regression
• ❌ Result: weak signal (~0.50 collapse)

Current Model

• 🌳 XGBoost

Why?

• Captures nonlinear patterns
• Handles interactions
• Produces useful probability spread

---

💡 Key Breakthrough

Before:

Probabilities ~0.50 → unusable

After XGBoost:

Probabilities ~0.15 → 0.86

✅ Enabled confidence-based trading decisions

---

🔄 System Evolution

Phase 1 — Pooled Model

• One model for all tickers
• Ranked signals globally

---

Phase 2 — Asset Filtering

• Removed weak contributors
• Improved stability

---

Phase 3 — Per-Ticker Models

• One model per asset
• Revealed asset-specific behavior

💡 Not all tickers behave the same

---

Phase 4 — Hybrid System (Current)

• Per-ticker models generate signals
• Signals compete globally
• Only the top opportunity per day is selected

---

📊 Signal Logic

Long  → prob_up ≥ 0.70
Short → prob_up ≤ 0.30
Else  → no trade

---

🏆 Ranking & Selection

Conviction

conviction = |prob_up - 0.50|

Selection Rule

top_n_per_day = 1

The system acts as a daily best-opportunity selector

---

📈 Backtesting

Trade Logic

• Long → next-day return
• Short → inverse return

Metrics

• Win rate
• Avg / median return
• Total return
• Coverage
• Per-ticker stats

---

📊 Current Performance

Configuration

• Thresholds: 0.70 / 0.30
• Top trades per day: 1

Results

• Coverage: ~5%
• Win rate: ~55%
• Avg return: ~0.24%
• Total return: ~90%+

---

📊 Visualizations & Results

The following plots were automatically generated from the backtest using generate_visualizations.py. They highlight the strategy's performance, signal quality, and per-asset contribution.

📊 Equity Curve

Download Interactive version: equity_curve_interactive.html

This chart shows the cumulative growth of the EdgeRank strategy over the test period. The annotation highlights the strong total return achieved with selective trading.

📉 Probability Distribution

This histogram demonstrates the key improvement from the XGBoost model: a wide, well-calibrated spread of predicted probabilities. Only high-conviction signals (≥ 0.60 for long, ≤ 0.40 for short) are acted upon.

📊 Performance by Ticker

Download Interactive version: per_ticker_performance_interactive.html

Breakdown of total strategy return by ticker. This clearly shows NVDA as a strong alpha driver while revealing weaker performers (e.g., GOOG).

📈 Trade Returns Distribution

Distribution of individual conviction-weighted trade returns. The positive skew supports the selective, high-conviction philosophy of the system.

---

All visualizations are regenerated automatically whenever the backtest is re-run.

---

🧠 Asset Insights

🟢 Strong

• NVDA (primary alpha driver)
• AAPL (consistent)
• SPY (stable)

🔴 Weak

• GOOG
• XLK
• NFLX

💡 Not all assets deserve equal participation

---

⚙️ System Behavior

Strengths

• Highly selective
• Robust over time
• Interpretable decisions

Observations

• Regime-dependent performance
• Strong in trending markets
• Flat periods exist

---

🧪 Experiment Tracking

• CSV logging system
• Tracks:
  • thresholds
  • ranking depth
  • performance

---

🧠 What This Demonstrates

• Technical features contain signal
• Nonlinear models outperform linear ones
• Selectivity > prediction volume
• Ranking is critical
• Specialization improves results

---

⚠️ Limitations

• No transaction costs
• No slippage
• Limited tickers
• No regime detection (yet)
• No macro inputs

---

🚀 Future Work

Short-Term

• Refine ticker set
• Test top_n = 2
• Add conviction filters

Medium-Term

• Regime detection
• Dynamic thresholds
• Risk controls

Long-Term

• Expand universe
• Ensemble models
• Optional macro features

---

🧭 Summary

EdgeRank is not just a model.

It is a decision system that identifies, ranks, and executes the highest-confidence opportunities in the market.

---

👤 Author

Andre — Computer Science @ UNM
Focus: AI / ML, Systems, Data Engineering