Deep-MKV-TS: Path-Dependent McKean–Vlasov Control for Financial Time Series Generation

Battle Count: 74

Highly relevant to quantitative trading. The paper directly addresses financial scenario generation for risk management, drawdown-risk budgeting, and conditional forecasting. The drawdown-risk decision experiment (Table 5) shows practical trading relevance: improved exposure capacity (1.99 vs 1.85) under a fixed 1% loss budget. The method preserves interpretability of the reference model while correcting path-dependent volatility features critical for trading decisions. Applications include stress testing, portfolio allocation, and generating training data for RL-based trading strategies. The conditional forecast evaluation on intraday equity-index futures (ES, NQ, YM) is directly applicable to trading workflows.

High complexity. Requires: (1) calibration of a path-dependent reference model with trend/activity memories, (2) derivation and implementation of the SMP forward-backward system, (3) MMD computation with custom kernel and feature maps, (4) GRU-based neural approximation of the backward adjoint signal, (5) ridge regression proxy for adjoint targets, (6) iterative training loop with forward simulation, adjoint computation via automatic differentiation, and neural projection. The algorithm involves multiple interacting components (reference model, control law, adjoint system, neural network, MMD loss). However, the architecture is relatively simple (1-layer GRU, 96 units) and the training procedure is clearly specified in Algorithm 1.

The paper provides detailed algorithm pseudocode (Algorithm 1), full training configuration (Table 6), evaluation metric definitions (Appendix B), and feature map definitions (Appendix A). However, no code repository is mentioned. The reference model parameters, GRU architecture (1-layer, 96-unit), and hyperparameters are specified. Data is from Databento (commercial). Four training seeds are used. The method builds on prior work [5,12,19] which are 'Work in Progress' or recent preprints, potentially limiting full reproducibility of the SMP derivation.