ASTRAEUS QUANT

THU-BDC LSTM Model Specifications

This system is designed specifically for the THU-BDC 2026 Big Data Challenge. It incorporates a state-of-the-art Attention-based LSTM (Long Short-Term Memory) Network that evaluates multi-channel financial timeseries inputs.

1. Mathematical Formulation

The core model projects stock price movements at a time horizon \(T+k\). The model optimizes the Mean Squared Error (MSE) augmented by a directional sign match penalty:

2. Key Features Utilized

• Technological Sentiment Factor: Crawled headlines analyzed via an embedding transformer.
• Volumetric Pressure Profile: Orderbook bid-ask depth imbalance calculated across 10 levels.
• Temporal Volatility Bounds: Rolling GARCH(1,1) estimates mapping conditional standard deviation.