Empirical Results

Measured results for the neural methods. The theory behind these methods is in the theory section.

Architecture-specific results

Each page focuses on finding the optimal network size/shape for each problem within a single architecture. Numbers are single-seed and indicative, not a rigorous hyperparameter optimum.

ArchitecturePageNetwork
Deep BSDEdeep_bsde.md\(Z\) (and \(\Gamma\)) nets
Deep HJB (DGM)dgm.mdvalue net
Neural operatoroperator.mdbranch/trunk (DeepONet) or spectral (FNO)

Cross-architecture results

PageContents
ResultsValidation, activation comparison, reference agreement
ComparisonTraining/execution time and error vs exact, split PDE/SOC
CoverageWhich architecture fits each exact-solution problem

Problems considered

ProblemClassExact solutionMath background
Black-Scholes European callPDEclosed formBSDE/DGM theory; PDE exact (Black-Scholes)
Merton log-utility portfolioSOCclosed formSOC exact (Merton portfolio)
Linear-quadratic regulatorSOCclosed form (Riccati)SOC exact (LQ regulator)
Jump linear-quadratic regulatorSOC (jump)closed form (Riccati/affine)SOC exact (LQ with jumps); jump theory
Avellaneda-Stoikov market makingSOCclosed form (matrix exponential)SOC exact (Avellaneda-Stoikov)
Merton jump portfolioSOC (jump)closed formSOC exact (Merton with deterministic jumps)
Merton jump log-normalSOC (jump)numerical rootSOC exact (Merton with log-normal jumps)
Jump-to-default portfolio (log utility, L1)SOC (jump)closed form (default-chain matrix exponential)Correlated price-default; operator theory
Jump-to-default portfolio (CARA L1)SOC (jump)no-default Merton; no-loss matrix exponential; L1 numericalCorrelated price-default
Jump-to-default portfolio (CARA L2/L3)SOC (jump)no closed form; DGM referenceCorrelated price-default

The Heston and Hawkes market-making extensions and unequal-\(k\) market making are not yet in neural_solver; they are the target of the neural-operator research plan and are documented in the neural operator theory. The log-utility L1 and CARA L1/L3 jump-to-default problems are implemented.

Selected CARA L3 result

The product-domain operator (Walsh on the default cube, MLP on prices) is benchmarked against a per-instance DGM on the S-dependent CARA L3 family: 0.77% normalized held-out RMSE with the Walsh layer and 1.51% without it. The operator trains in 0.51 s; the DGM reference is 1.37 s per instance. The multi-source deep BSDE matches the DGM within 0.25-1.41% per instance. The detailed table is in the source page operator.md.