FORECASTING STOCK PRICES IN NAIROBI STOCK EXCHANGE USING MACHINE LEARNING AND TIMESERIES

Abstract

The volatility, structural inefficiencies, and external shocks, are common in emerging markets such as Nairobi Securities Exchange (NSE). These make predictive modelling challenging in the markets hence stock price forecasting in financial markets has been a critical challenge particularly in emerging markets. This work helps fill gaps in forecasting techniques by creating and testing hybrid models, which combine machine learning and time series methods. The study compared two models, the Autoregressive Integrated Moving Average (ARIMA) and Artificial Neural Network (ANN), it enhances ANNs with Ordinary Differential Equation (ODE)-based features, which can capture multi-scale behavior, and it also augments ARIMA with GARCH model to model volatility clustering. The study utilizes closing prices of Safaricom PLC (2015-2020), a strong stock in the NSE, on a daily basis, to test these frameworks. The ARIMA models were based on linear tendencies, and the ANN model used the feature engineering based on the lagged stock prices, the derivatives to identify the short-term fluctuations, and the integrals to identify the long-term tendencies. The hybrid ARIMA-GARCH accounted volatility persistence. Model performance was measured by the Mean Squared Error (MSE), Mean Absolute Error (MAE), and the Root Mean Squared Error (RMSE). The key findings made included the fact that the ODE-enhanced ANN was better than the standalone ARIMA and the simple ANN models in terms of the decrease of the MSE by 64.7 percent and the decrease of the MAE by 71. The results of the ARIMA-GARCH hybrid were impressive as they reduced RMSE by 17.4 percent and confirmed the applicability of the volatility modeling in the emerging markets. The superior performance of ANN shows that it was able to capture non-linear dynamics and to respond to the extreme fluctuations of the price, and volatility persistence (β = 0.999) in GARCH showed high volatility persistence in NSE that made NSE susceptible to long-term instabilities. The researchers have suggested application of hybrid models among the practitioners of the financial markets as they are able to capture the characteristics of the domain. The research also emphasizes the need for innovative feature engineering in enhancing risk mitigation and trading strategies. To policymakers, this research recommends that it is necessary to reduce market concentration and increase data transparency so as to level the current volatility witnessed in emerging economies. Further areas of research potential would be a combination of other data sets, such as news data, sentiments, economic indexes, among others, to increase the accuracy of the predictive behavior.