A1014
Title: Sentiment analysis in an LSTM deep learning approach for forecasting volume in fractional trading
Authors: Andrew Rosenswie - HTW University of Applied Sciences Berlin (Germany)
Christina Erlwein-Sayer - University of Applied Sciences HTW Berlin (Germany) [presenting]
Alla Petukhina - HTW Berlin (Germany)
Sakir Kepezkaya - HTW University of Applied Sciences Berlin (Germany)
Abstract: Fractional trading has emerged as an opportunity for investors to invest in fractions of stocks. It gains popularity since trades in large companies and high-valued stocks can be realized even for small budgets and a bigger investment community. Risks of fractional trading for sellers and brokerages lie mostly in forecasting the number of fractional trades over a short period since those trades potentially need to be balanced out. News sentiment analysis and machine learning approaches are combined for time series modelling to find the best-suited models to forecast trading volume in fractional trading. A long-short-term memory neural network (LSTM) is built to forecast volume. The results are further investigated with measures of feature importance to find the most important explanatory variables. News sentiment analysis is combined with LSTM and most relevant features are discovered. The model and case studies focus on four major companies in different sectors to identify the most relevant features. The model is built in two steps: a cluster analysis finds company clusters with similar trading behaviors. In the second step, the volume is modelled through news-sentiment-based LSTM. Preliminary results show clear indications of the main explanatory variables, including news sentiment. Volume in fractional trading is modeled, and reliable features are found for forecasting volume. Findings show that the LSTM performance is mostly improved when news sentiment is added.
