A COMPARATIVE ANALYSIS OF TIME SERIES MODELS FOR PREDICTING THE S&P SL 20 INDEX OF THE COLOMBO STOCK EXCHANGE (CSE)

Abstract

The performance of stock markets is influenced by various factors, and understanding these dynamics is crucial for investors and policymakers. This research centers on Sri Lanka's capital market, with particular attention to the Colombo Stock Exchange (CSE), and analyzes the influence of the Standard & Poor's Sri Lanka 20 (S&P SL 20) index, which represents the top 20 leading companies listed on the CSE. Specifically, the primary objective of this research is to compare the effectiveness of traditional time series models with machine learning and deep learning models in predicting the S&P SL 20 index. These models, developed using computerized programs, are evaluated based on their predictive performance within the context of the CSE. The study will use daily S&P SL 20 stock index data obtained from the CSE data library enclosing the period 2010 to 2018. This methodology compares Autoregressive Integrated Moving Average (ARIMA), which is a traditional time series model and Long Short-Term Memory (LSTM), which is a recurrent neural network model. In this research, the Python language will be employed for analysis. The ARIMA and LSTM models are evaluated using three performance metrics: MAE, MAPE, and RMSE. ARIMA slightly outperforms LSTM in MAE (233.96 vs. 249.37) and RMSE (269.57 vs. 269.86), which essentially indicates better overall accuracy in absolute and squared error terms. However, LSTM achieves a marginally lower MAPE (6.96% vs. 7.04%), showing fewer relative percentage errors. All in all, both models offer similar performances, with minor differences depending on the metric. Both ARIMA and LSTM show strengths in predicting the S&P SL 20 Index. ARIMA excels in minimizing absolute errors (MAE), ideal for linear trends. LSTM's lower MAPE highlights its ability to capture nonlinear patterns. With similar RMSE values, both handle overall errors well. ARIMA's constant predictions in some periods reveal its limitations with limited or weak trend data. The choice depends on the forecasting goal: ARIMA for linear trends and LSTM for complex patterns.