Climate-Dyna Deep Hedging for XVAs: Model-Based Reinforcement Learning, Residual Climate HVA, and Hedge-Instrument Discovery

Battle Count: 55

Highly relevant for quantitative risk management and hedging desks. The paper directly addresses XVA valuation, deep hedging with trading frictions, model-based RL for data-poor regimes, and hedge-instrument discovery. The residual accounting framework (crediting inherited hedges before overlay optimization) is directly applicable to trading desk P&L attribution. The Riccati anchor + RL correction architecture is practical for production hedging systems. The EU ETS calibration makes it immediately relevant to carbon-market trading desks. However, it is more focused on risk management and hedging than on alpha generation or algorithmic execution.

Very high complexity. Requires: (1) exact Riccati backward recursion with Möbius iteration for scalar case, (2) constrained structural-residual world model ensemble with masked-softmax regime rows and absorbing default, (3) paired world simulation with common shocks, (4) prioritized Dyna rollouts with importance correction, (5) risk-sensitive entropic policy gradient with tilted weighting, (6) cross-fitted critic-based Q-learning with fold exclusion, (7) independent deployment gating with UCB on reference and stressed-liquidity models, (8) Shapley attribution for instrument value, (9) pathwise accounting verification. Multiple interacting components with careful constraint enforcement.

The paper uses public data (EEX auctions, ECB rates, CISS, EBA PD/LGD, EU ETS data) and provides detailed algorithmic descriptions including Riccati recursions, Dyna planning, and gating procedures. However, no code repository is mentioned. The semi-synthetic environment construction (paired worlds, pre-MSR baseline) involves scenario-specific choices. Exact dynamic programming is used as evaluation oracle. The 20 held-out datasets and specific hyperparameters (25 transitions, 8 folds, 4 world models, 128 samples) are described but full implementation details for the neural residual policy and world model architecture are not fully specified.