AI-driven portfolio optimization is the use of machine learning to decide how much of a portfolio to hold in each asset. It extends classical portfolio optimization, which follows modern portfolio theory and needs forecasts of return, risk and covariance for every asset. Those forecasts are noisy, and small errors in them can produce extreme and unstable weights.
Machine learning is applied in three main ways. Models can forecast the inputs, such as expected returns or volatility. Clustering methods can group related assets before risk is shared out, as in hierarchical risk parity. And reinforcement learning can be trained to pick weights directly, with trading costs included in its objective. The main weakness is overfitting: financial history is short and changes character, so a model that fits the past closely may fail on new data. Results are judged on data held back from training, after costs.