Augurus

Augurus is a financial market prediction system utilizing a Transformer architecture for next-token prediction and RL fine-tuning.

Market movement prediction using Augurus as indicator in OpenBackTest

Architecture

It employs a hybrid architecture based on a GPT-style model. This approach embeds and quantizes mixed continuous and discrete inputs into tokens, which are then processed using multi-head self-attention.

Project Structure

• /training: Python-based training pipeline (PyTorch) and export it to ONNX format.
• /inference: C#-based API for real-time market predictions (ONNX).

Training Pipeline

1. Data Preprocessing

Converts raw Binance OHLCV CSV data into tokenized features and calculates rolling volatility.

python process_data.py

• Input: data/BTCUSDT_5m_...csv
• Output: processed/tokenized_data.csv, processed/quantizer.pkl

2. Phase 1: Next-Token Prediction (Pre-training)

Trains the Transformer to predict the next market "token" (price movement bin). This builds the model's understanding of market dynamics.

python train.py

• Config: SEQ_LEN=64, BATCH_SIZE=128
• Output: processed/model_epoch_N.pt

3. Phase 2: RL Fine-tuning (Actor-Critic)

Fine-tunes the pre-trained model using Reinforcement Learning to optimize for actual trading PnL (Profit and Loss). (Not necesary for good results)

python rl_train.py

• Requires: A checkpoint from Phase 1 (e.g., model_epoch_5.pt).
• Output: processed/model_rl_epoch_N.pt

4. Export to ONNX

Converts the trained PyTorch model into ONNX format for high-performance inference in the C# Augurus.Api.

python export_onnx.py

• Output: ../inference/financial_transformer.onnx, ../inference/quantizer.pkl

---

Hardware Notes

• CUDA: Recommended for train.py and rl_train.py.
• Memory: Reducing SEQ_LEN to 64 allows for larger BATCH_SIZE (e.g., 128) on consumer GPUs.