-
Notifications
You must be signed in to change notification settings - Fork 2
Roadmap
Maxime Grenu edited this page Feb 18, 2026
·
1 revision
| Sprint | Theme | Key Goals | Status |
|---|---|---|---|
| Sprint 1 | Foundation | Core ML models, paper trading, security hardening | ✅ Done |
| Sprint 2 | Reinforcement Learning | PPO/DQN agents, live data feeds, Telegram integration | 🔄 In Progress |
| Sprint 3 | Multi-Exchange | Portfolio-level risk, multi-asset support | 📋 Planned |
| Sprint 4 | Optimization | Walk-forward validation, hyperparameter search | 📋 Planned |
| Sprint 5 | Production Hardening | Full observability, kill-switch v2, compliance | 📋 Planned |
Goal: Build the core trading infrastructure and validate the ML pipeline end-to-end.
Delivered:
- RandomForest, NeuralNetwork, Transformer, Ensemble models
- Paper trading harness with simulated order execution
- Backtesting framework
- Risk manager (position sizing, stop-loss, max drawdown)
- Kill-switch (hard halt on drawdown threshold)
- HashiCorp Vault integration for secrets
- Telegram bot alerts
- Prometheus + Grafana monitoring stack
- Initial paper trading results: Sharpe 0.69, drawdown 0.009%, win rate 70%
Goal: Add RL-based agents that learn from market interaction, not just historical patterns.
Planned:
- PPO (Proximal Policy Optimization) agent for continuous action spaces
- DQN (Deep Q-Network) agent for discrete buy/sell/hold decisions
- Live WebSocket data feed integration (Binance)
- Replay buffer and online learning pipeline
- Enhanced Telegram reporting (equity curve images, daily summaries)
- Walk-forward validation on Sprint 1 models
Goal: Expand beyond Binance to support portfolio-level trading across multiple exchanges.
Planned:
- Kraken and Coinbase Pro integration via CCXT
- Portfolio-level position tracking (cross-exchange)
- Correlation-aware risk manager (avoid over-concentration in correlated assets)
- Multi-asset strategy: BTC, ETH, SOL pairs
- Unified orderbook aggregation
Goal: Systematically find better parameters and validate strategies out-of-sample.
Planned:
- Hyperparameter optimization (Optuna / Ray Tune)
- Walk-forward validation framework (train/test splits with rolling windows)
- Regime detection (trending vs. ranging market classifier)
- Ensemble weight optimization (meta-learner training)
- Factor analysis (feature importance, SHAP values)
Goal: Make the system safe enough (in concept) for extended paper trading and potential future live use.
Planned:
- Kill-switch v2: multi-layer (per-asset, per-strategy, global)
- Vault HSM integration (hardware key storage)
- Full audit log (every trade decision with feature snapshot)
- Regulatory compliance skeleton (position limits, reporting)
- Chaos testing (simulate exchange outages, data gaps, model failures)
- Canary deployment pattern for strategy updates
⚠️ Reminder: ELVIS remains a research project. Even after Sprint 5, live deployment requires independent risk review, legal clearance, and extensive additional validation.