Option pricing using artificial neural networks (ANN) model while relaxing the assumption of constant volatility still remains a challenge. The conventional practice for pure ANN models has been to either model volatility using the very ANN model and have the model output fed as an input to the ANN option pricing model, or to make allowances for a large number of lags directly as inputs to the option pricing model with the belief that the ability of ANN to incorporate flexibility and redundancy creates a more robust model. This has been done in spite of a well-known fact-that financial time series data harbors a set of characteristics such as volatility clustering, leptokurtosis and leverage effects-features that ANNs in their pure forms have proved inadequate in capturing. Consequently, this study sought to follow the conventional methods employed by other studies and developed two pure ANN option pricing models-one with constant volatility and the other while violating the assumption of constant volatility with an aim of establishing whether significant differences exist in the outputs of the two models. The intraday data for the AAPL stock option for the period between December 2016 and March 2017 with 56,238 data points was used in validating the developed models. Results indicate that the ANN model (with varying volatility) makes better predictions than the model with constant volatility. However, the difference between the performance of the two models is not significant at 0.05 level of significance.

In the absence of a well-defined input selection technique associated with the pure ANN models, Option pricing using pure ANN models while relaxing the assumption of constant volatility remains a challenge. The conservative drill espoused has been to make allowances for a large number of input lags with the confidence that the ability of ANN to integrate suppleness and redundancy generates a more robust model. This is to say that the nonexistence of input selection criteria notwithstanding, the models have been developed without due consideration to the effect that the choice of input selection technique would have on model complexity, learning difficulty and performance measures. In this study, we deviate from the conventional techniques applied by the pure ANN option price models and adopt the hybrid model in which the volatility component is handled using some celebrated time series models, with specificity to the ANN-GJR-GARCH model-a hybrid of the ANN and a time series hybrid. The hybrid ANN option pricing model is then framed and tested with the forecasts of the ANN-GJR-GARCH model as a volatility input alongside two other inputstime to maturity and moneyness. Finally, we compare the performance of the hybrid model developed with that of a pure ANN model. Results indicate that the hybrid model outperforms the pure ANN model not only in forecasting but also in the training time and model complexity.