BOT 28P: An Open-Source Multi-Model Ensemble Trading System Combining WFO-Trained Deep Learning with LLM Veto

BOT_28P: System Summary

BOT_28P is an open-source, multi-model ensemble trading system designed for quantitative finance. It introduces a hybrid "open-core" architecture that combines public code with private model weights, aiming to resolve the industry tension between the need for transparency and the protection of competitive advantages.

FULL CODE 

Core Architecture and Innovation

The system utilizes a unique three-tier pipeline to process trades:

• MLM12 Ensemble: At the core are three 12-layer hybrid neural networks (MLM12) specifically trained for BTC, ETH, and SOL. These models feature a sophisticated architecture including convolutional layers for pattern extraction, LSTM layers for temporal memory, and attention mechanisms for importance weighting.

• Dual LLM Veto: This is the system's primary innovation. Instead of using Large Language Models to generate trade ideas—which often leads to hallucinations—BOT_28P uses DeepSeek and Qwen as "rejectors." A trade is only executed if both models agree the risk is below a specific threshold. This "conservative ensemble" requires a consensus for approval, effectively acting as a safeguard against false positives.

• Risk and Execution: The final layer applies data-driven parameters discovered through Walk-Forward Optimization (WFO). This ensures the strategy requires no human discretionary rules and adapts to the specific volatility profiles of different assets.

Key Technical Components

• Walk-Forward Optimization (WFO): The system uses a rolling window training and validation procedure to prevent overfitting. This process discovered 29 distinct parameters per asset, such as specific ATR-based take-profit and stop-loss levels.

• Asymmetric Error Costs: The logic assumes that an LLM blocking a good trade (opportunity loss) is an acceptable, low-cost error, whereas an LLM allowing a bad trade (capital loss) is a high-cost error that must be mitigated through strict veto thresholds.

• Open-Source Implementation: The entire decision logic, execution engine, and parameter sets are public on GitHub, while only the specific trained model weights remain proprietary.

Experimental Results

In testing conducted on Google Colab using the NVIDIA NeMo framework, BOT_28P achieved a 100% pass rate across 15 production scenarios, including valid trades and edge cases. Initial trade execution statistics showed an average risk/reward ratio of 2.30:1 and an estimated success rate of 70%.

Future Development

While the current version operates in a simulated environment, future work is prioritized for live exchange deployment, the integration of additional LLMs like Claude or GPT-4, and the expansion of the asset pool to include tokens like ADA and MATIC. The system stands as a paradigm shift in transparent algorithmic trading, leveraging AI strengths while strictly controlling for their known weaknesses.