Improving Forecast Accuracy With Hybrid Time Series Forecasting And Machine Learning Methods: A Case Study Of A Pharmaceutical Company

Abstract

Pharmaceutical industry is instrumental to the well-being of people around the world as it deals with the discovery, development, production and marketing of drugs and medications, which are one of life’s necessities. Recently, there is an emerging phenomenon that some hospitals must suppress their operation due to the deficiency of medical supply, and thus affecting patients who are in the need for treatment. One of the factors contributing to the situation is the incapability of providing these commodities on time from pharmaceutical companies, which results from their inaccurate forecasting. This thesis investigates the application of machine learning algorithms in time series forecasting, specifically focusing on Support Vector Regression (SVR) to improve forecast accuracy. It evaluates the performance of five time series models, including Moving Average, ARIMA, Exponential Smoothing, Holt’s Trend, and Holt’s Winters, and three types of SVR, including Linear kernel, Polynomial kernel, and Radial Basic Function (RBF). The evaluation is based on the Mean Absolute Percentage Error (MAPE) metric, which measures the accuracy of the forecasts. The study then proposes hybrid models that combine SVR with traditional time series models via ensemble methods and compares their performance with the individual models. Finally, a case study of a pharmaceutical company is performed to assess the impacts of these methods. The results demonstrate that the hybrid ensemble method achieves lower MAPE values compared to the individual models. These findings offer promising insights and highlight the potential of the proposed hybrid models in improving forecast accuracy.