Kraken Meta-Agent Trading System
This project introduces a regime-aware autonomous trading architecture that shifts the focus from predictive modeling to adaptive portfolio allocation. Instead of attempting to forecast price movements, the system emphasizes robustness through dynamic regime detection and controlled exposure. At its core, the architecture separates the execution layer from the model evolution layer. The execution layer operates as a deterministic trading agent, handling market data ingestion, signal processing, risk constraints, and order execution via Kraken integration. It focuses strictly on allocation and exposure management rather than prediction. Above it, a meta-agent system continuously refines and evolves trading models using feedback from market performance. This recursive loop enables continuous improvement without destabilizing live execution. The system integrates regime scoring based on momentum, volatility, drawdown, relative strength, and persistence, allowing capital allocation to adapt across changing market conditions. Portfolio construction follows a hierarchical structure: capital is first allocated across asset classes using regime scores, then distributed within each class using model-driven signals, including GRU-based forecasts and technical indicators. The result is not a conventional trading bot, but a self-improving financial system designed for stability, adaptability, and long-term survival in complex markets.
Category tags:"1. Application of Technology: 4.5 / 5 Justification: Very impressive engineering. Implementing a custom GRU (Gated Recurrent Unit) forward-pass runtime in raw numpy (gru_runtime.py) so the execution agent doesn't need to load PyTorch into memory is an incredibly sophisticated architectural flex. 2. Presentation: 4 / 5 Justification: Solid presentation. The concept of a "recursive loop enabling continuous improvement without destabilizing live execution" is a great narrative. The repository is well-documented with appropriate architecture guides. 3. Business Value: 4.5 / 5 Justification: High value. In traditional quantitative finance, portfolio allocation across asset classes based on market regime (volatility, momentum trend) is much more important for long-term survival than attempting to snipe 1% arbitrage trades. This project correctly frames the problem as an allocation and risk exposure problem rather than a prediction game. 4. Originality: 4 / 5 Justification: Moving away from LLM-driven prediction (which is what 90% of hackathon entrants do) and instead building a classical Machine Learning (GRU) regime-aware allocation system is a breath of fresh air. ⚖️ Pros & Cons Pros: Deep Learning Runtime: Utilizing GRU neural networks for forecasting sequences brings a higher tier of machine-learning sophistication than standard moving-average bots. Allocation over Prediction: Emphasizing Portfolio Risk Allocation (how much capital goes into an asset based on the current regime) rather than just binary Buy/Sell signals is a very mature, production-ready mindset. Separation of Concerns: Hard-separating the execution layer (which runs tight, safe, local Python loops) from the evolving "meta-agent" layer prevents catastrophic logic failures from bleeding into live market execution. Cons: Complexity to Audit: Because the signals are driven by GRU hidden-state neural networks rather than transparent LLM prompts or explicit algorithmic indicators, it becomes a "black box," making it harder to debug why the bot made a decision during a flash crash. Missing On-Chain Layer: Unlike STRIKER or ZKSentinel, this project seems entirely focused on the Kraken CLI execution side and Neural Network logic, missing the ERC-8004 cryptographic verification/trust requirements that provide security guarantees to outside investors."
Sanem Avcil
For Innovators & Creators