ARTEMIS: A Neuro-Symbolic Framework for Economically Constrained Market Dynamics

Battle Count: 65

Highly relevant to quantitative trading. Directly addresses three core challenges in deploying deep learning for trading: (1) interpretability for regulatory compliance via symbolic bottleneck producing closed-form alpha factors, (2) economic plausibility through no-arbitrage constraints and Sharpe ratio bounds preventing spurious signals, (3) continuous-time modeling for irregular high-frequency data. Achieves state-of-the-art directional accuracy (64.96% on crash regime DSLOB, 96.0% on Time-IMM), which is the primary metric for trading strategies. Includes conformal prediction for risk-aware position sizing and Kelly criterion portfolio optimization. Ablation study demonstrates that physics-informed losses are critical for directional signal (removing them drops accuracy below random). However, underperformance on volatility prediction (Optiver) and computational overhead limit immediate production deployment.

Very high implementation complexity. Requires expertise in: (1) Laplace Neural Operator with rational function kernel parameterization in Laplace domain, (2) Neural SDE simulation with Euler-Maruyama scheme and reparameterization trick for backpropagation, (3) Physics-informed losses requiring automatic differentiation through PDE residuals (Feynman-Kac), (4) Market price of risk computation with Cholesky decomposition for SPD diffusion matrices, (5) Differentiable symbolic regression with Gumbel-Softmax and two-phase teacher-student training, (6) Adaptive conformal prediction with rolling windows, (7) Differentiable convex optimization for Kelly criterion portfolio allocation. Multiple interacting loss terms require careful hyperparameter tuning. Streaming data loaders and mixed-precision training needed for large datasets. The framework integrates 4+ distinct research areas (neural operators, SDEs, PINNs, symbolic regression, conformal prediction) into a single computational graph.

Code and data stated to be made publicly available upon publication. Four datasets used: Jane Street (Kaggle), Optiver (Kaggle), Time-IMM (public), and DSLOB (synthetic, proprietary). Detailed algorithm pseudocode (Algorithm 1) and mathematical derivations (Appendix A) provided. Hyperparameters partially specified (loss weights λ1-λ4, Sharpe threshold κ, Gumbel temperature τ, risk aversion γ). However, exact hyperparameter values for all experiments are not fully disclosed, and DSLOB generation parameters are described but the synthetic data generation code availability is uncertain. Five independent runs with different seeds reported with Wilcoxon signed-rank test (p<0.01).