Relevance
9/10
Directly addresses continuous-time portfolio allocation with risk-sensitive objectives and benchmark-relative performance. The fractional Kelly decomposition provides interpretable allocation rules. The LQG structure and affine policies are directly applicable to factor-based equity allocation. The ergodic (infinite-horizon) formulation is particularly relevant for institutional asset management. The proof-of-concept uses real U.S. equity ETF data with a six-factor model. The main limitation for direct trading application is the partly model-based nature and absence of transaction costs.
Implementation Complexity
8/10
High complexity due to: (1) free energy-entropy duality reformulation requiring Girsanov change of measure, (2) LQG stochastic differential game with two players, (3) Riccati equation system for value function, (4) continuous-time martingale-based actor-critic updates, (5) multiple parametrized function approximations (actors, critics, q-functions), (6) RLS for gradient critic learning, (7) central finite differences for TD targets, (8) preconditioning and stabilization of actor updates. The analytical structure simplifies some aspects but the overall framework requires deep expertise in stochastic control, game theory, and RL.
Reproducibility
3/5
The paper provides detailed algorithms (Algorithm 1 and 2), explicit parametrizations, and uses publicly available data (Kenneth French's data library for factors, CRSP for ETF returns). However, CRSP requires a subscription. The implementation is in R but no code repository is explicitly linked. The analytical model parameters are fully specified, and the proof-of-concept is well-described. The partly model-based nature means some coefficients are assumed known.