TradeAlgo is a modular quantitative trading research framework designed to evaluate trading signals derived from a combination of:

News sentiment analysis (NLP-based) Macroeconomic event surprises Technical indicators (ATR, EMA) Multi-factor signal aggregation

The system integrates these signals into a unified scoring model (S-score) and evaluates performance through a realistic event-driven backtesting engine that includes spreads, slippage, and dynamic position sizing.

This project demonstrates end-to-end capability in:

Data engineering Signal processing Financial modeling Backtesting system design NLP integration into quantitative workflows System Architecture

The pipeline is structured into four core components:

1. Data Layer Historical FX OHLCV data (EURUSD 1H) News dataset with metadata (author credibility, engagement signals) Economic event calendar (actual vs consensus)
2. Signal Generation Layer Sentiment extraction using transformer-based NLP model Event surprise quantification (actual vs forecast deviation) Technical confirmation signals (ATR, trend filters)
3. Aggregation Layer Rolling sentiment window with credibility-weighted scoring Novelty adjustment based on unique information flow Multi-factor signal fusion into a single S-score
4. Execution & Backtesting Engine Event-driven simulation Spread + slippage modeling ATR-based dynamic stop-loss / take-profit Risk-based position sizing Equity curve tracking + performance metrics Signal Model

The core decision variable is:

S-score = f(sentiment, event impact, novelty, technical confirmation)

Where:

Sentiment reflects aggregated NLP sentiment over a rolling window Event impact measures macroeconomic surprise magnitude Novelty estimates information freshness / crowding Technical confirmation validates market structure alignment

Trade decisions are derived from:

BUY if S exceeds positive threshold SELL if S falls below negative threshold HOLD otherwise Backtesting Engine Features

The backtester simulates realistic execution conditions:

Bid/ask spread modeling Slippage approximation Stop-loss and take-profit execution Time-based trade exit horizon Position sizing based on account risk % Equity curve generation Performance statistics: Total PnL Win rate Average trade return Maximum drawdown Risk Model

Position sizing is dynamically computed using:

Fixed fractional risk model (default: 0.5% per trade) ATR-based stop-loss distance Pip-value normalization for FX instruments

This ensures:

consistent risk exposure across volatility regimes scalable position sizing based on account equity Technologies Used Python 3.10+ Pandas / NumPy HuggingFace Transformers (RoBERTa sentiment model) Time-series simulation logic Custom event-driven backtesting framework How It Works Load historical FX price data Stream news and macro events chronologically Compute rolling sentiment score Align macroeconomic surprises with price timeline Generate S-score from multi-factor fusion Execute trades in simulated environment Track equity curve and performance metrics

📈 Example Output Total PnL: 5833.10 Win rate: 75% Max drawdown: -1.0% Trades executed: 12

Note: Results are dependent on dataset size, signal thresholds, and backtest configuration.

Limitations

This project is a research-grade prototype, not a production trading system.

Current limitations include:

Limited sample size in backtests Simple threshold-based decision logic No walk-forward validation No parameter optimization framework No execution latency modeling No live market integration Future Improvements

Planned enhancements:

Signal calibration using quantile-based decisioning Walk-forward validation pipeline Benchmark strategies (MA crossover, random baseline) Sharpe ratio + statistical significance testing Regime detection (trend vs mean reversion) Probabilistic position sizing model Live data ingestion (broker API integration)