Fortune: A Financial Quantitative Analysis System Driven by Hybrid Human-Machine Intelligence

Fortune is a financial quantitative analysis system integrating large language model (LLM) reasoning and machine learning (ML) prediction, focusing on the Hong Kong stock market, cryptocurrency, and gold markets. Its core innovation lies in the hybrid human-machine intelligence architecture, which achieves fully automated trading signal generation and delivery through multi-time dimension prediction and cross-validation strategies, with commercial potential. The project is open-sourced on GitHub and maintained by wonglaitung.

Project Basic Information

• Original Author/Maintainer: wonglaitung
• Source Platform: GitHub
• Original Link: https://github.com/wonglaitung/fortune
• Release Date: 2026-06-14

Core Philosophy

Traditional quantitative systems rely on single indicators or models and struggle to handle market sentiment and complex events. Fortune proposes the hybrid human-machine intelligence philosophy, combining LLM's ability to understand unstructured information with ML's precise prediction. It maintains objectivity while flexibly adapting to market contexts, focusing on real-time monitoring and analysis of the Hong Kong stock market, cryptocurrency, and gold markets.

Hybrid Human-Machine Intelligence Architecture

• LLM Reasoning Layer: Processes unstructured information such as macro events, policy changes, and market sentiment
• ML Prediction Layer: Predicts rise/fall probability based on 1023 technical features
• Cross-Validation Mechanism: Multi-time dimension signal resonance triggers recommendations, reducing false positive rates

Multi-Time Dimension Prediction System

Prediction Cycle	Accuracy	Features	Usage
1-day	51.49%	High noise, for reference only	Intraday trading reference
5-day	65.86%	Trend confirmation, auxiliary decision	Weekly position decision
20-day	81.22%	Most reliable, main decision basis	Monthly investment direction

Three-cycle cross-validation supports short, medium, and long-term trading.

Core Trading Mode Performance

The system identifies 8 trading signal modes (1/0 represents rise/fall), with core modes as follows:

Mode	Description	20-day Accuracy	Strategy Recommendation
101	False breakout (1 rise, 5 fall, 20 rise)	87.32%	Best long entry opportunity, short-term pullback followed by mid-term rise
111	Consistent bullish	86.26%	Add positions after trend confirmation
001	Downward continuation (1 fall,5 fall,20 rise)	81.05%	Buy signal after decline
000	Consistent bearish	79.80%	Reduce positions or short

Four Core Trading Rules

1. False breakout long (signal 101): 87.32% accuracy
2. Consistent bullish buy (signal 111): 86.26% accuracy
3. Downward continuation long (signal 001):81.05% accuracy
4. Consistent bearish short (signal 000):79.80% accuracy

Machine Learning Model

Adopts the CatBoost gradient boosting algorithm, with integrated feature dimensions including: technical indicators (MA, RSI, etc.), price patterns, fundamental data, market sentiment, capital flow, interest rate features, volatility models, market state, calendar effects, and network features.

Top 10 Feature Importance

1. Volatility_30pct (30% quantile volatility)
2. MA250_Slope (250-day moving average slope)
3. Volatility_30d (30-day volatility)
4. BB_Width_MA60 (Bollinger Band width 60-day average)
5. net_cohesion_HSI_Regime_Duration (network cohesion × Hang Seng Index regime duration)
6. Volatility_70pct (70% quantile volatility)
7. Distance_Support_120d (distance to 120-day support level)
8. net_cohesion_per_GARCH_Conditional_Vol (network cohesion × GARCH conditional volatility)
9. Stock_Price_Stability_Score (stock price stability score)
10. 60d_Trend_HSI_Return_60d (60-day trend × Hang Seng Index 60-day return)

Walk-forward Validation Results

Performance under conditions including 57 Hong Kong stocks and 12-fold cross-validation:

Indicator	Value	Industry Standard	Evaluation
Comprehensive Score	90/100	-	Excellent
Average Sharpe Ratio	5.33	>0.5	Excellent
Average Max Drawdown	-1.04%	<-20%	Excellent
Average Accuracy	55.04%	>50%	Qualified
Average IC	0.205	>0.05	Excellent
Average Rank IC	0.231	>0.05	Excellent
Average Return	+5.08%	>0%	Positive

The system achieves full automation via GitHub Actions scheduled workflows, covering:

• Data collection
• Feature calculation
• Model prediction
• Trading signal generation and email delivery

Daily auto-generated outputs:

• Hong Kong stock trading recommendation reports
• Prediction accuracy evaluation
• Monthly/quarterly/annual statistical summaries

No manual intervention is required, ensuring timely capture of trading opportunities.

Fortune's value lies in demonstrating the application potential of hybrid human-machine intelligence in finance:

1. Complementary Advantages: LLM's unstructured information understanding + ML's pattern recognition, achieving 1+1>2
2. Interpretability: Provides clear trading rules and pattern explanations, avoiding black-box models
3. Transparency: All predictions are validated via backtesting, with no hidden failure results
4. Focus: Optimized for Hong Kong stocks (southbound capital, Hang Seng Index correlation, etc.)

For developers, Fortune provides an excellent reference architecture and implementation example for intelligent quantitative systems.