Host: The Japanese Society for Artificial Intelligence
Name : The 31st Annual Conference of the Japanese Society for Artificial Intelligence, 2017
Number : 31
Location : [in Japanese]
Date : May 23, 2017 - May 26, 2017
This paper examines the empirical properties of trading volume and the predictability of trading volume and absolute return with the Long Short Term Memory(LSTM). With the analysis on the S&P500 index/firms daily data, two properties are discovered: i) Trading volume has a long memory ii) Trading volume is essentially related to the price return distribution. The latter part aims to exploit long memory of financial markets for the prediction. The LSTM model, an architecture designed for modelling long memory is constructed for the prediction of trading volume and absolute return with their past values. The prediction with the model mainly achieved three results: iii) Trading volume is highly predictable with its past values. iv) The LSTM model overwhelms the performance of the GARCH(1,1) with input of temporally distant past values and without addition of variables. v) The contribution of trading volume to the prediction of absolute return is insignificant.