Back to Projects
LangGraphConvNeXtGAFPhi-3.5pgvectorPython
Hybrid-Adaptive Quant Trading System
Combines GAF pattern recognition with ConvNeXt-Tiny for market regime and direction prediction. LLM confidence calibration filters low-confidence signals. Vector memory stores historical patterns for adaptive strategy refinement.
Phi-3.5-Mini & ConvNeXt-Tiny
Core Technology
Gramian Angular Fields
Pattern Recognition
78% BSC Model
Confidence Threshold
31K Tokens
Context Window
pgvector with HNSW
Vector Database
Python, LangGraph
Backend
# Methodology
• GAF Pattern Recognition: Converts time series into 2D images using Gramian Angular Fields. Captures price momentum, angle fields, and volatility patterns in 3-channel format for CNN processing.
• ConvNeXt-Tiny Neural Network: CNN backbone predicts market regime (trending/ranging/volatile), price direction, and volatility in parallel. Uses depthwise convolutions and gradient checkpointing for efficiency.
• LLM-as-a-Judge Calibration: LLM refines confidence estimates using Binary Symmetric Channel model. Triggers ensemble re-processing when confidence drops below 78%. Filters out weak signals before execution.
• Memory & Learning: pgvector stores prediction history in embedding space. Similarity search finds similar market conditions from past periods. Updates strategy parameters based on realized outcomes using temporal difference learning.
• ConvNeXt-Tiny Neural Network: CNN backbone predicts market regime (trending/ranging/volatile), price direction, and volatility in parallel. Uses depthwise convolutions and gradient checkpointing for efficiency.
• LLM-as-a-Judge Calibration: LLM refines confidence estimates using Binary Symmetric Channel model. Triggers ensemble re-processing when confidence drops below 78%. Filters out weak signals before execution.
• Memory & Learning: pgvector stores prediction history in embedding space. Similarity search finds similar market conditions from past periods. Updates strategy parameters based on realized outcomes using temporal difference learning.
Core Model
# Data Source
Real-time market data via yfinance (Futures, Crypto, ETFs)