Overview
Built a walk-forward, ranking-based ML system for ETF execution targeting VOO and VXUS, improving mean monthly return from ~1.28% to ~1.46% and Sharpe-like ratio from ~0.32 to ~0.35 versus SIP baselines across 150+ months of historical data.
Problem
Standard SIP (Systematic Investment Plan) strategies allocate capital on a fixed schedule regardless of market conditions. The goal was to build an agentic system that dynamically decides when and how much to invest based on ML-predicted confidence scores.
Approach
- Walk-forward validation across 150+ months to prevent data leakage
- Ranking-based execution — only invest when model confidence exceeds threshold
- Automated feature engineering on macroeconomic and price-based signals
- Containerized pipeline for scheduled monthly inference
- GenAI reporting — auto-generated investment decision summaries
Results
| Metric | SIP Baseline | ML System | |--------|-------------|-----------| | Mean monthly return | ~1.28% | ~1.46% | | Sharpe-like ratio | ~0.32 | ~0.35 | | Months backtested | 150+ | 150+ |
Tech Stack
Python scikit-learn pandas Docker GenAI LLM
Key Learnings
Walk-forward validation is non-negotiable in financial ML — even a single look-ahead bias invalidates the entire backtest. Confidence-based execution (only trading when the model is sufficiently certain) was the single biggest driver of Sharpe improvement.