Transfer Learning (Il)liquidity

Battle Count: 83

Highly relevant for quantitative trading in options markets. The RND is fundamental for pricing, hedging, and constructing optimal portfolios. The ability to recover RND from as few as three option quotes is particularly valuable for illiquid markets, tail risk assessment, and strategies that rely on implied distributions. The convexity guarantee ensures no-arbitrage compliance, which is critical for trading systems. However, the method is more suited for research and risk management than high-frequency trading due to the computational requirements of the deep learning pipeline.

The Deep-LSE architecture requires custom implementation combining affine layers, Log-Sum-Exp operations, positive recursion weights via softplus, and multi-layer composition. The transfer learning pipeline adds complexity with two-phase training and optimal stopping criteria based on KL divergence. Theoretical guarantees (convexity, universal approximation, consistency) require careful parameter constraint enforcement. The architecture is not available in standard ML libraries and would need to be built from scratch using frameworks like PyTorch or TensorFlow. The sieve estimation framework and consistency proofs add mathematical complexity for proper implementation.

The paper provides detailed mathematical formulations, simulation parameters (Tables 1, 4, 5, 6), and algorithm pseudocode (Algorithm 1). However, no code repository is mentioned. The SPX data used is publicly available (January 2015 and 2016). The Deep-LSE architecture is fully specified with equations, and theoretical proofs are provided. Reproduction would require implementing the custom architecture from scratch.