Embedded Files
Grant: 411PED/2020, PN-III-P2-2.1-PED-2019-2271
Contracting authority: UEFISCDI
Grant value: 600.000 RON
Coordinating institution: University of Bucharest, Human Language Technologies Research Center
Implementation period: 01.11.2020 - 31.10.2022
Abstract:
Stock price prediction has been an ever-going challenge for economists but also for machine learning scientists. They have all long searched for means to find those specific patterns that trigger the dynamic behavior of the market and predict its next move. The usual trading indicators of a company are the number and value of transactions, the number of shares, the minimum, average and maximum prices, the open and close price. The prediction commonly targets the close price. Taking into account its values for a number of days back, the task is to forecast its value for the next day. The analysis of such historical time series data has been tackled in literature through different computational means, starting from the classical ARIMA models to the new generation of deep learning techniques. The project team has also pursued the close price prediction task through both ARIMA and deep learning on real-world, numerical long-term data from 25 companies enlisted at the Bucharest Stock Exchange. However, the indicators that set off the behavior of the stock market may be out of this box of historical trading. Therefore, the next big challenge would be to also use within the prediction models the possible sentiment of the traders with respect to a company, when deciding for a major buy or sell action. Such a decision generates a cascade of the corresponding trading operation within the other market players and eventually marks an increase/decrease in the close price trend of its stocks. An optimistic or pessimistic sentiment can be triggered only by what appears in the press regarding that company. For the Romanian stock market, the channel is the newspaper Bursa. As such, besides the numerical trading data from BVB, the text articles related to the watched companies can be additionally analyzed for an optimistic/pessimistic sentiment and the dual information be simultaneously used to enhance the prediction accuracy of the learning models regarding the future trend.
Obtained results
The INTEREST project aimed for the analysis of the influence of the sentiment produced by the financial specialized media articles on the trading actors in the stock exchange environment, regarding the decision to buy or sell the shares of listed companies.
The result of the project was an artificial intelligence framework (deep learning), which aimed both at the analysis of the sentiment from texts and at the prediction of time series. First of all, texts from the specialized website bursa.ro regarding companies listed on the Bucharest Stock Exchange were analyzed. Positive/negative, pessimistic/optimistic feelings, as well as joy/surprise/anger/disgust/fear/sadness emotions were extracted from these. These sentiments were included in the predictive analysis of a company's share price fluctuation.
The conclusion is that the media-induced sentiment is correlated with the buy-sell decision, and artificial intelligence can provide a support tool for the financial field.
Page updated
Google Sites
Report abuse