An investigation into the use of machine learning techniques for forecasting inventory stock

• Main Author: Samuel, Neethu
• Format: Others
• Language: en
• Published: 2021
• Online Access: https://hdl.handle.net/10539/31267

A research report submitted to the Faculty of Engineering and the Built Environment, University of the Witwatersrand, in partial fulfilment of the requirements for the degree of Master of Science in Engineering, 2020 === Demand forecasting is the field of study that aims to predict customer demand. In the past, demand forecasting was achieved using traditional stochastic methods like the Holtz Winters method, ARIMA and moving averages. Machine learning techniques are techniques that can capture the characteristics of data more efficiently. Thus; machine learning techniques were explored for the purpose of demand forecasting in this research report. The dataset used for this research is Kaggle’s Historical Product Demand. This dataset consists of various product demand categories arranged in monthly logs from 2011 to 2017. There are 2172 product categories in the original dataset. After processing, there are 1803 product categories. The dataset is essentially a time series dataset; so, machine learning and statistical methods can be applied to it. The machine learning techniques utilized for the research are Artificial Neural Networks (ANNs) and Support Vector Regression (SVR). The results of the forecasting using ANNs was compared to SVR; which was then validated against the forecast obtained from an ARIMA method. It was discovered that for series with no trend and seasonality and only irregular and/or cyclical behaviour, the SVR Gaussian model is the clear performer in 92% of the product series. The remaining 8% have the ANN traincgf algorithm as the model with the smallest MAD on the overall dataset. For series with slight trend and no seasonality, the Gaussian SVR kernel outperforms the other kernels for 85% of the products. The ANN trainlm algorithm performs the best for 9% of the products, followed by the SVR polynomial kernel for the remaining 6%. For series with clear/heavy trend and no seasonality, the ANN trainbfg algorithm outperforms the other training algorithms === CK2021