Portfolio Optimization, Tested Live
A portfolio optimization AI for a portfolio manager, on Nasdaq and BIST data. In a three-month live trial, its portfolio groups returned more than 60% on average.
The problem.
Financial models fail in a specific way: they perform well on the period they were fitted to and badly on the one that matters. A portfolio built on such a model inherits the failure, usually at the worst moment.
A portfolio manager needs more than a forecast. They need an allocation, and a way to know how far to trust it before money moves.
The system.
We built a portfolio optimization AI as an MVP. It works on historical Nasdaq and BIST data, at daily and at high frequency.
It forecasts each asset and wraps every forecast in a conformal prediction interval, which states how wide the error can be. The forecasting models build on our own published research in time-series prediction.
It then optimizes the allocation along the efficient frontier, with those intervals as inputs, so an asset the models are unsure of carries less weight. Finally it stress-tests the result against market regimes the models were not fitted to: drawdowns, correlation breakdowns, and liquidity squeezes.
In operation.
In a three-month live trial, the portfolio groups it built returned more than 60% on average. Three months is a short window, and past returns are no promise of future ones. The intervals and the stress tests exist for exactly that reason.
The engagement closed at the MVP stage, and the system did not go on to production. All three parts were built and tried live; production would have added the monitoring and controls a fund needs to run it every day.
Ready to build your AI capability?
We accept a limited number of research partnerships per quarter to ensure high-fidelity delivery.
Book a Discovery CallNot ready for a call? contact@whitehand.ai