Hybrid Forecasting of Stock Prices in Pakistan’s Sugar Sector: Integrating Supervised and Unsupervised Machine Learning under Market Uncertainty

Abstract

The task of forecasting stock trading values has become more complex, especially for the turbulent and fragmented sectors such as the sugar industry in Pakistan. The paper sets forth the use of Statistical approaches and classifies a mix of models involving supervised (ANN and LSTM) and unsupervised machine learning methods (K-Means and PCA) for forecasting. Data comprises daily stock prices covering the period from 2012 to 2023 of eight sugar companies listed on the stock exchange. Valuation models were constructed by completing the quantitative method within its context to compute various error measures such as RMSE and MAPE. As Hybrid models perform better when compared with ARIMA and GARCH models, especially when utilizing them regarding volatility predictions. The study noted differences among models; the model within the augmented architecture proved to work exceptionally well in the high volatility period. On the other hand, the ANN model was found to perform well on stable companies. The study presents the synergies that shaped the adoption of adaptive hybrid support mechanisms in highly uncertain and unstable market conditions. In this case, we might imagine abandoning the Efficient Market Hypothesis and moving towards the Adaptive Market approach in its place. In the same context, this study also offers better methods of forecasting, becoming useful to stakeholders within the finance and public spheres, as well as strategists concerned only with one sector of the economy. The end also does not disregard the application of these models solely to a specific sector, but also the need for such models to be available in a format that can be used as and when required. Of all predictive models examined, the LSTM-GARCH model showed the lowest RMSE of 8.95, which is far better than ARIMA’s 17.75 and therefore can be considered good progress.