rosemarycox5334-debug/AlphaMaster

What it solves

AlphaMaster is a quantitative trading factor mining center that automates the discovery of interpretable factor formulas from historical market data (Parquet or MT5 K-lines). It eliminates the manual trial-and-error process of creating trading strategies by using machine learning to search for the most effective mathematical formulas for predicting market movements.

How it works

The system uses deep reinforcement learning to search through a space of features and operators to find optimal formulas. These formulas are represented as token sequences and executed via a custom StackVM. The workflow is divided into three main stages:

1. Training: The RL agent searches for formulas and selects the best ones based on validation set performance.
2. Backtesting: The system simulates trading using a continuous position sizing logic (tanh(factor)) to evaluate performance metrics like Sharpe and Sortino ratios.
3. Real-time Analysis: The system monitors live data sources (such as MT5 or OKX) and recalculates signals based on the closed K-line of the current period to avoid signal flicker.

Who it’s for

Quantitative traders and developers who want to automate the factor mining process and move from historical training to real-time signal generation and deployment.

Highlights

• End-to-end Pipeline: Integrates training, backtesting, and real-time signal analysis in a single Web UI.
• Interpretable Formulas: Unlike "black box" models, it searches for explicit mathematical formulas that can be executed by a StackVM.
• Consistent Signal Logic: Uses the same signal calculation logic across training, backtesting, and realtime analysis to prevent look-ahead bias.
• External Notifications: Supports Feishu (Lark) Webhooks for alerts when trading signal directions change.