Aiming at the problem of prediction bias in GARCH volatility model, the prediction bias of GARCH model is corrected based on improved recurrent neural network. This paper constructs an option pricing model with predictable volatility in an effort to reduce the estimation error of the model and attain more accurate volatility estimation. The experimental results indicate that the model can significantly update the prediction accuracy of GARCH model. Based on this, the stochastic differential and martingale methods are used to get the option pricing formula under the risk neutral condition. To certain degree, it eliminate the shortcoming of the traditional artificial designed volatility model, which can only estimate the parameters by using the option market price. The numerical results show that the precision of the system is ideal, and it is found that the B-S formula for 50ETF stock option pricing is often less than the option pricing in this paper.