Deep Learning and Elicitability for McKean-Vlasov FBSDEs With Common Noise

Battle Count: 75

The paper is primarily a numerical methods contribution for solving MV-FBSDEs, which underpin mean-field game equilibria relevant to systemic risk and macro-finance. While not directly about trading strategies, the methods apply to: (1) systemic risk monitoring in inter-bank networks, (2) mean-field equilibrium computation for large-population trading models, (3) tail-risk and quantile-based risk measures via the elicitability framework. The quantile-interaction extension is particularly relevant for risk management. However, the paper does not address execution, alpha generation, or direct trading applications.

High complexity due to: (1) multi-step Picard iteration scheme with inner and outer loops, (2) training multiple neural networks (RNN for S, hybrid FF-RNN for Y, RNNs for Z and Z0) at each iteration, (3) elicitability-based loss construction requiring careful score function design, (4) soft-update damping mechanism, (5) Euler-Maruyama discretization of forward SDEs, (6) time-dependent weighting in backward loss, (7) coordination between multiple trained networks for post-training sampling. Requires strong background in SDEs, mean-field theory, and deep learning.

Code is publicly available on GitHub. The paper provides detailed algorithm pseudocode (Algorithms 1-6), specific hyperparameters (batch size 8192, 2000 backprop iterations, AdamW lr=0.0005, 101 timesteps, 50000 paths, 20 Picard iterations, damping delta=0.5), network architecture details (18 hidden units, SiLU activation, 4 residual blocks), and parameter values for all three experiments. However, random seeds and exact training curves are not fully specified.