Multi-agent AI platform for stock research
Combining quant analysis, ML signals, news intelligence and sentiment into a unified trading verdict.
BenderEdge is a portfolio and research project. It is NOT financial advice.
- All analysis is generated by AI models and automated systems
- Results may be inaccurate, outdated, or misleading
- Never make investment decisions based solely on this tool
- Past backtest performance does not guarantee future results
- The creator assumes no responsibility for any financial losses
Use at your own risk.
BenderEdge is a multi-agent AI system that analyzes stocks using 5 specialized agents working in parallel. Each agent produces a BUY / SELL / HOLD vote with confidence score. A portfolio agent aggregates all signals using weighted voting to produce the final verdict.
Most stock analysis tools give you a single signal from a single model. BenderEdge is different:
- 7 agents with different methodologies vote independently
- Disagreement is a signal — low agreement = lower confidence
- ML model trains on-the-fly for each ticker, no stale pre-trained weights
- Sector-aware XGBoost — different hyperparameters per sector
- Multi-horizon signals — 1-7 day, 1-3 month, 3+ month predictions
- SHAP explainability — understand what drove each decision
- BenderScanner — scan entire indices for opportunities
- Accuracy tracking — validate predictions against real price movements
- Fully local — Ollama LLM, no API costs, no data sent to third parties
- Real-time streaming — watch each agent work as it happens
User Input (Ticker / Index)
│
▼
FastAPI Orchestrator (LangChain)
│
├─── Researcher Agent ──► NewsAPI + LLM analysis
├─── Quant Agent ──► RSI, MACD, BB, SMA, ATR risk mgmt
├─── Sentiment Agent ──► Headline tone scoring
├─── BenderQuant Agent ──► XGBoost on-the-fly + SHAP
├─── Insider Agent ──► SEC EDGAR Form 4
├─── Macro Agent ──► FRED API — rates, inflation, GDP
├─── Earnings Agent ──► EPS history + surprise
│
▼
Portfolio Agent (Weighted Voting)
│
▼
BUY / SELL / HOLD + Confidence Score
│
├──► SSE Streaming → Next.js Dashboard
└──► SQLite → History + Accuracy Tracking
Fetches live financial news via NewsAPI. Identifies the most market-moving headlines, extracts risk signals and catalysts, and votes with confidence.
Output: Summary · Spotlight headline · Risk · Catalyst · Vote
Computes 11 technical indicators from price history using yfinance.
| Indicator | Signal |
|---|---|
| RSI (14) | Oversold / Overbought |
| MACD | Bullish / Bearish crossover |
| Bollinger Bands | Price position relative to bands |
| SMA 20 / 50 | Trend direction |
| Support / Resistance | Key price levels |
| Volume spike | Unusual activity |
Analyzes news headline tone and produces a numerical sentiment score with directional label.
Output: Score (-10 → +10) · Bullish / Neutral / Bearish · Vote
The most sophisticated agent. Powered by XGBoost, trained on-the-fly using 2 years of OHLCV data for the requested ticker.
Multi-horizon signals:
- 1-7 day short-term prediction
- 1-3 month mid-term prediction
- 3+ month long-term prediction
Sector-aware training: Each sector uses optimized XGBoost hyperparameters — Technology, Financial, Energy, Healthcare, and 7 more sectors each have tuned parameters for better accuracy.
SHAP Explainability: Every decision comes with feature importance — see exactly which indicator drove the model's verdict.
Model quality metrics:
- 5-fold cross-validation score
- Test set accuracy
Backtest (2 years):
- Total return · Sharpe ratio · Max drawdown · CAGR
- Equity curve · Recent trade log
Fundamentals:
- P/E ratio · EPS · Market Cap · Beta
- 52-week high/low · Sector · AI company detection
Powered by BenderQuant — a standalone XGBoost financial classification project.
Aggregates all agent votes using a weighted system designed to prioritize quantitative signals.
| Agent | Weight | Rationale |
|---|---|---|
| Quant | 30% | Objective, rule-based |
| BenderQuant ML | 25% | Data-driven, ticker-specific |
| Sentiment | 15% | News-driven momentum |
| Researcher | 10% | Contextual intelligence |
| Earnings | 10% | Fundamental performance |
| Macro | 5% | Economic environment |
| Insider | 5% | Smart money signals |
confidence_score = agent_agreement × signal_strength
Scan entire market indices automatically. BenderScanner runs all 5 agents on each ticker and ranks opportunities by signal strength.
Supported indices:
- S&P 500 (top 20)
- NASDAQ 100 (top 20)
- BIST 50
- BIST 100
Accuracy Tracker: Every analysis is saved to SQLite. After 7 days, the system automatically validates predictions against real price movements and calculates:
- Overall accuracy %
- Accuracy by verdict (BUY / SELL / HOLD)
- Average return per signal type
Ticker: TSLA
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
PORTFOLIO VERDICT: SELL
Confidence: High (0.79)
Weighted Score: -0.862
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Agent Votes:
Researcher → HOLD (75%) mixed signals
Quant → SELL (75%) bearish MACD, below SMA
Sentiment → SELL (75%) UK sales drop -45%
ML → SELL (97%) 0/5 BUY signals last 5 days
Agreement: 75% | BUY: 0 SELL: 3 HOLD: 1
Multi-horizon:
1-7 day → SELL 97%
1-3 month → SELL 83%
3+ month → SELL 99%
SHAP Top Features:
RSI +1.28 (BUY pressure)
Volatility -0.48 (SELL pressure)
MACD +0.37 (BUY pressure)
BenderQuant Backtest (2Y):
Return: +340% | Sharpe: 2.1 | Max DD: -9.2%
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
| Layer | Technology | Purpose |
|---|---|---|
| Frontend | Next.js 14 + Tailwind + Recharts | Dashboard UI |
| Backend | FastAPI + Python | API + orchestration |
| ML | XGBoost + scikit-learn + SHAP | On-the-fly model training |
| LLM | Ollama (Qwen2.5:7b) + LangChain | Agent reasoning |
| Data | yfinance + NewsAPI | Price + news data |
| Streaming | Server-Sent Events (SSE) | Real-time agent output |
| Database | SQLite | Analysis history + validation |
- ✅ Real-time agent streaming — each agent result appears as it completes
- ✅ Tab-based dashboard — Overview / Quant / ML Model / Research
- ✅ Weighted multi-agent voting with quantified confidence scoring
- ✅ On-the-fly XGBoost training — no stale pre-trained weights
- ✅ Sector-aware hyperparameters — 11 sectors with optimized params
- ✅ Multi-horizon ML signals — 1-7 day / 1-3 month / 3+ month
- ✅ SHAP feature importance — explainable AI decisions
- ✅ Full backtesting — Sharpe, Max Drawdown, CAGR, equity curve
- ✅ Fundamental analysis — P/E, EPS, Market Cap, Beta
- ✅ BenderScanner — scan S&P500, NASDAQ100, BIST50, BIST100
- ✅ Accuracy tracking — validate predictions vs real price movements
- ✅ Analysis history — SQLite persistence across sessions
- ✅ Global market support — NYSE, NASDAQ, BIST, LSE, Frankfurt, Tokyo
- ✅ Spotlight headlines with Risk / Catalyst extraction
- ✅ Fully local LLM — no API costs, no data leakage
- ✅ 7 specialized agents — Quant, ML, Sentiment, Insider, Macro, Earnings, Research
- ✅ ATR-based risk management — stop loss, take profit, position sizing
- ✅ Walk-forward backtesting — out-of-sample validation
- ✅ SMOTE class balancing — handles imbalanced training data
- ✅ Sidebar dashboard — clean navigation between analysis sections
- ✅ Landing page — quick-start with recent tickers
# 1. Clone
git clone https://github.com/yusufbender/BenderEdge.git
cd BenderEdge
# 2. Backend setup
python -m venv venv
venv\Scripts\activate # Windows
source venv/bin/activate # macOS/Linux
pip install -r requirements.txt
# 3. Environment
echo "NEWS_API_KEY=your_key_here" > backend/.env
# 4. Pull local LLM
ollama pull qwen2.5:7b
# 5. Start backend
cd backend
uvicorn main:app --reload
# 6. Start frontend (new terminal)
cd ../frontend
npm install
npm run devOpen http://localhost:3000
Scanner: http://localhost:3000/scanner
| Market | Suffix | Example |
|---|---|---|
| NYSE / NASDAQ | — | AAPL TSLA NVDA |
| Borsa Istanbul | .IS |
THYAO.IS GARAN.IS |
| London | .L |
HSBA.L |
| Frankfurt | .DE |
SAP.DE |
| Tokyo | .T |
7203.T |
| Hong Kong | .HK |
0700.HK |
- 5-agent voting architecture
- FastAPI + SSE streaming backend
- Next.js tab-based dashboard
- BenderQuant XGBoost integration
- On-the-fly model training
- Backtest + equity curve
- Fundamental data
- Global market support
- 1-7 day short-term signal
- 1-3 month mid-term signal
- 3+ month long-term signal
- Per-horizon XGBoost models
- Sector-based XGBoost hyperparameter optimization
- 11 sector profiles (Technology, Financial, Energy...)
- Automatic sector detection via yfinance
- SHAP feature importance
- Top 5 features driving each decision
- Visual bar chart in ML Model tab
- Multi-index scanner (SP500, NASDAQ100, BIST50, BIST100)
- SQLite analysis history
- 7-day prediction validation
- Accuracy tracker by verdict type
- Insider Trading Agent (SEC EDGAR Form 4)
- Macro Agent (FRED API)
- Earnings Agent (EPS beat/miss history)
- Landing page with hero section
- Sidebar navigation
- Navbar with ticker + new analysis button
- ATR-based risk management
- Walk-forward backtesting
- SMOTE class balancing
- Model determinism (random_state=42)
- Rate limiting
- Redis cache
- MLflow experiment tracking
- LangGraph agent orchestration
- User authentication
- Alert system
- Deploy on Railway + Vercel
- Docker Compose setup
BenderEdge/
├── backend/
│ ├── agents/
│ │ ├── researcher.py # News + LLM analysis
│ │ ├── quant.py # Technical indicators
│ │ ├── sentiment.py # Headline tone scoring
│ │ ├── ml_agent.py # XGBoost + SHAP + multi-horizon
│ │ └── portfolio.py # Weighted voting
│ ├── routers/
│ │ ├── analysis.py # SSE + REST endpoints
│ │ └── scanner.py # Market scanner endpoints
│ ├── database.py # SQLite — history + validation
│ ├── models/ # Saved model artifacts
│ └── main.py # FastAPI app
└── frontend/
└── app/
├── page.tsx # Main dashboard
└── scanner/
└── page.tsx # BenderScanner
BenderQuant
XGBoost financial classification model. Trained on multi-ticker datasets with feature engineering, SMOTE oversampling, and GridSearch hyperparameter tuning. Powers the ML agent in BenderEdge.
Yusuf — AI / ML Engineer
Building production-grade AI systems, LLM architectures, and intelligent financial tools.
⭐ If you find this project useful or interesting, consider giving it a star.
This project is for educational and portfolio purposes only. Not financial advice.