Portfolio optimization based on neural networks sensitivities from assets dynamics respect common drivers

We present a framework for modeling asset and portfolio dynamics, incorporating this information into optimization. define drivers dynamics their optimal selection. For framework, we introduce the Commonality Principle, providing solution selection of as common drivers. Portfolio constituent are modeled by Partial Differential Equations, solutions approximated with neural networks. Sensitivities respect to obtained via Automatic Adjoint Differentiation. Information on is incorporated sensitivities constituents embedded space drivers, distance matrix in called Sensitivity used solve convex optimization diversification. The sensitivity measures similarity projections vector formed drivers’ returns optimize diversification both idiosyncratic systematic risks while adding directionality future behavior dynamics. optimization, perform hierarchical clustering matrix. tree recursive bisection obtain weights. To best author’s knowledge, first time that sensitivities’ networks have been Secondly, problem incorporate these sensitivities. Thirdly, public listed variables can be maximum means reach over-performance many experiments all other out-of-sample methods different markets real datasets. also include recipe methodology increase performance even further, tackle main issues management such regimes, non-stationarity, overfitting, bias.