Case Study
AI-Driven Portfolio Management System
Designed a deep learning portfolio management system translating neural network outputs directly into deployable portfolio weights, with a custom risk-adjusted cost function and multi-regime backtesting.
Challenge
DCM wanted to replace static portfolio allocation with a data-driven, adaptive framework — one where a model trained on historical return data could dynamically adjust asset weights across market regimes and produce execution-ready allocations without manual intervention.
Solution
Helios architected a custom neural network framework inspired by Deep Portfolio Theory (Heaton, Polson, Witte 2017), where the network structure mirrors portfolio composition — each node representing an asset and inter-layer weights corresponding to allocation decisions. The model was trained end-to-end on historical return data with a custom cost function incorporating risk-adjusted return maximisation, volatility penalisation, drawdown sensitivity, and capital efficiency constraints. A portfolio construction and rebalancing engine supported threshold-based, time-based, and hybrid rebalancing strategies with turnover minimisation and drift correction. Backtesting and simulation frameworks validated performance and weight stability across multiple market regimes.
Technical Responsibilities
- Designed an advanced portfolio management system using deep learning for dynamic asset allocation
- Architected a custom neural network based on Deep Portfolio Theory (Heaton, Polson, Witte 2017), where network weights correspond directly to portfolio allocation decisions
- Implemented end-to-end training pipelines on historical return data to learn optimal portfolio weights across varying market regimes
- Designed a custom cost function incorporating risk-adjusted return maximisation, volatility penalisation, drawdown sensitivity, and capital efficiency constraints
- Translated model outputs directly into deployable portfolio weights, bridging research models and execution-ready allocations
- Built a portfolio construction and rebalancing engine supporting threshold-based, time-based, and hybrid rebalancing strategies
- Designed portfolio drift detection and correction systems ensuring consistent alignment with target exposures and risk profiles
- Built data pipelines for ingestion, preprocessing, and feature engineering of historical market data
- Developed backtesting and simulation frameworks to evaluate strategy performance, weight stability, and sensitivity across market regimes
- Designed a modular research-to-production pipeline enabling seamless transition from model training to allocation to execution
Outcomes
- Systematic, model-driven portfolio allocation framework adapting dynamically to market conditions
- Custom neural network architecture producing execution-ready portfolio weights directly from return data
- Performance validated across multiple market regimes via backtesting and simulation
- Modular pipeline bridging research, model training, and live allocation with minimal friction