Researchers Solve Multi-Period Mean-DCVaR With RNN

Researchers present a new method to solve discrete-time multi-period portfolio optimization with an explicit tail-risk constraint, using a `recurrent neural network` to approximate the optimal `precommitment policy` for a mean-`DCVaR` objective. The paper frames the constraint as the excess of Conditional Value-at-Risk over expected terminal wealth and addresses the time-inconsistent nature of the precommitment formulation. Validation occurs in a classical complete-market model and an application to multi-period portfolio allocation in (re)insurance, capturing long-horizon liability dynamics.

The core technical move is to replace dynamic programming and pathwise dual methods with a learning-based control approximation that directly parameterizes the policy with a recurrent neural network and optimizes by stochastic gradient methods. Key elements include:

• •A formalization of `DCVaR` as a global, path-dependent constraint that admits an exact-penalty reformulation for feasibility control
• •A policy class based on recurrent architectures that preserves memory of past states and actions, making it suitable for path-dependent constraints
• •Numerical validation in a complete-market model and an insurance liability setting showing feasibility enforcement and improved scalability versus grid-based DP

The contribution sits at the intersection of quantitative finance and data-driven control. Classic methods for multi-period tail-risk constraints either collapse into dynamic programming, which does not scale, or use duality and scenario-based approximations that struggle with path dependence. By using a learnable recurrent policy and exact-penalty constraints, the authors offer a practical route for high-dimensional portfolios and long-horizon liabilities where state space explosion and path dependence block traditional solvers. This aligns with a broader trend of replacing intractable stochastic control problems with parameterized policies trained end-to-end.

The method scales to higher-dimensional state dynamics and enforces a transparent feasibility notion via penalties, but performance depends on optimization stability, penalty tuning, and the representational capacity of the chosen recurrent architecture. The paper validates the approach in stylized complete-market and insurance models; real-world market frictions, transaction costs, and model misspecification remain open challenges.

Look for released code, empirical tests on historical market data with transaction costs, and extensions to alternative tail measures or robust formulations. Adoption in actuarial practice or asset-liability management would be an early real-world signal of practical utility.