| Summary: | 碩士 === 國立中山大學 === 財務管理學系研究所 === 106 === Many studies exhibit that LSTM can outperform other traditional econometric and machine learning models. However, there are still somewhere insufficient we would like to make up to the literature. For the first time, we include more meaningful input variables and separate the noise from the input variables of the LSTM model, attempting to improve the predictability of the models. In addition, in order to reveal the knowledge that the LSTM model learn, we assign the stocks into bullish group and bearish group based on the predictive values of the stock prices, and then exhibit the descriptive statistics for each two group. Based on our empirical results, first of all, we acquire more precise predictive results after extending and denoising the input variables of the LSTM model. Furthermore, we notice the existence of the unlearned characteristics in the testing sets weakens the predictability of the LSTM model. Finally, we find our LSTM model captures the patterns of commonly known market anomalies about price contrarian, overreaction and underreaction. However, the market anomalies referred above are not quite significant among large cap stocks.
|
|---|
