Research

European Option Pricing with Deep Learning-Based CEV and SLV Models

基于深度学习的 CEV 与 SLV 模型欧式期权定价

Accepted · MATHEMATICA NUMERICA SINICA · In Chinese
With Meng Cai (蔡猛)

We study deep learning-based European option pricing under CEV and stochastic local volatility models, comparing five specifications within a unified risk-neutral framework. I built the pricing and calibration pipeline using Monte Carlo simulation, antithetic variance reduction, truncated Euler discretization, and two-path unbiased stochastic gradients. On 354,999 filtered 50ETF option quotes, neural CEV reduced out-of-sample relative MAE by approximately 21% versus Black–Scholes, while the more flexible neural SLV specification did not consistently improve pricing accuracy.

Revisiting the Idiosyncratic Volatility Puzzle in China: Evidence from Deep Learning Models

Working paper · Apr. 2025 – Jan. 2026 · SSRN

I re-examined the idiosyncratic volatility puzzle using 25 years of equity data and 22 volatility forecasting models to investigate whether the negative IVOL–return relation reflects systematic forecast bias. TFT reduced mean forecast error by 14.5% relative to the Martingale benchmark. Portfolio sorts and Fama–MacBeth regressions showed that the pricing relation depended on the forecasting specification; removing the top 3% of overpredicted observations rendered the Martingale-based IVOL coefficient insignificant. Decomposing returns into intraday and overnight components further explained the differences across models.

Comparative evaluation of asset pricing factor models

Research Assistant · Empirical asset pricing · Related journal article

As a research assistant, I supported an empirical comparison of asset pricing factor models in the Chinese A-share market. I prepared market and accounting data, replicated 105 anomaly characteristics, and implemented time-series and cross-sectional model comparisons in Python and Stata. The analysis combined portfolio construction, GRS tests, bootstrap resampling, and robustness checks to evaluate the models’ explanatory power.