Exploiting univariate and multivariate LSTM models for stock price
forecast: The case of Colombo Stock Exchange
Dhanushki Mapitigama* 1 , Shanaka Munasinghe 1 , Shehan Perera 1 and Namal Karunarathna 1
1
Department of Computer Engineering, Faculty of Engineering, University of Peradeniya, Sri
*E-mail: [email protected]
Abstract: With rapid economic growth, most entities focus on investing the money to obtain greater revenue.
The stock market is the most accessible investment opportunity for almost everyone and therefore forecasting
the trends of a stock market is of significant interest to researchers. Although highly volatile and chaotic nature
of the trends makes such forecasting challenging, it can be done by considering all relevant information that
affects the upward and the downward trends. All previous work on forecasting the trends of stock markets are
performed for stable stock exchanges such as Newyork and Shanghai. However, such forecasting models would
not capture stock exchanges which show high disruptive behaviours like the Colombo Stock Exchange (CSE).
Further, stock exchanges such as CSE do not have trades on certain indexes on certain days making the data
incomplete. In our work, we propose a forecasting model for stock exchanges with such disruptive behaviours
and incomplete data. Our effort focuses on experimenting with univariate and multivariate Long Short Term
Memory (LSTM) models to find the best-optimized model with the best set of features to predict stock prices
and forecast the trend of indices in CSE, which is can be used for an automated trading platform. The multivariate
LSTM model with the day of the week and trend for the past days along with past stock prices as features
gave 54% accuracy with around 0.04 Root Mean Square Error (RMSE) score which is better than the compared
baseline models.
Keywords - Stock market, Colombo Stock Exchange, LSTM, Time Series Analysis, Quantitative Analysis
20