The performance and error analysis of LSTM model combined with various GARCH models in stock forecasting

Volatility is a measure of the asset return rate's estimated level of uncertainty and may be used to assess the riskiness of financial assets. We use the market capitalization of 200 stocks representing Korea as the primary analytical aim and evaluate the accuracy between them by analyzing the impacts of different hybrid models' hybrid neural networks, which are based on the returns of the KOSPI 200 stock index. By measuring the effectiveness of these models using four dissimilarity measures, we contrasted the performance of hybrid models that combine a single neural network and a single GARCH type model with that of hybrid neural networks that combine multiple GARCH models (MAE, MSE, HMAE, and HMSE). They are applied to anticipate the KOSPI 200 index data's actual volatility. Among these, hybrid neural networks that integrate more than one GARCH-type model have much better forecasting performance than neural network models that mix two or more or more GARCH-type models. GW-LSTM makes the least accurate forecast. We note that the hybrid model combining the three GARCH models shows a minor increase in predicting ability based on merging two and three models.