Generative Pricing of Basket Options via Signature-Conditioned Mixture Density Networks

Battle Count: 76

Highly relevant for quantitative trading desks dealing with multi-asset derivatives. The train-once, price-anywhere paradigm with ~3.4ms inference latency makes it suitable for real-time pricing in production trading systems. The ability to handle time-varying rates, local volatility, and varying basket compositions without retraining addresses key operational challenges in derivatives pricing. However, it is limited to European options and does not directly address trading strategy development, execution, or portfolio construction. The surrogate model approach is particularly valuable for risk management teams needing rapid scenario analysis.

Moderate to high complexity. Requires: (1) Monte Carlo simulation infrastructure for generating 20M training samples under GBM with CIR-simulated parameters, (2) Path signature computation (truncated at level 5) from rough path theory, (3) Custom MDN architecture with constrained outputs (softmax for weights, softplus*tanh^2 for std dev), (4) LogSumExp stabilization for numerical training, (5) Cholesky decomposition for correlation structures, (6) Integration of learned density against payoff functions for pricing. The signature computation and MDN training pipeline require specialized knowledge in both rough path theory and deep learning. However, the architecture itself is relatively standard (6-layer MLP with mixture output head).

The paper provides detailed architecture specifications (6 hidden layers with neuron counts 320, 256, 256, 192, 128, 80), training hyperparameters (AdamW optimizer, learning rate 0.01, batch size 100,000), data generation algorithms (Algorithm 1 and 2), CIR model parameters, local volatility functional form, and hardware specifications (Apple M1 Pro, 16 GB RAM). However, no code repository is provided, and specific numerical KL divergence and pricing error values are only shown graphically in figures rather than tabulated. The training data is synthetically generated via Monte Carlo simulation, making it reproducible in principle but requiring significant computational resources (20 million samples).