A Comparative Study on Statistical and Machine Learning Forecasting Methods for an FMCG Company

Demand forecasting has been an area of study among scholars and businessmen ever since the start of the industrial revolution and has only gained focus in recent years with the advancements in AI. Accurate forecasts are no longer a luxury, but a necessity to have for effective decisions made in planning production and marketing. Many aspects of the business depend on demand, and this is particularly true for the Fast-Moving Consumer Goods industry where the high volume and demand volatility poses a challenge for planners to generate accurate forecasts as consumer demand complexity rises. Inaccurate demand forecasts lead to multiple issues such as high holding costs on excess inventory, shortages on certain SKUs in the market leading to sales loss and a significant impact on both top line and bottom line for the business. Researchers have attempted to look at the performance of statistical time series models in comparison to machine learning methods to evaluate their robustness, computational time and power. In this paper, a comparative study was conducted using statistical and machine learning techniques to generate an accurate forecast using shipment data of an FMCG company. Naïve method was used as a benchmark to evaluate performance of other forecasting techniques, and was compared to exponential smoothing, ARIMA, KNN, Facebook Prophet and LSTM using past 3 years shipments. Methodology followed was CRISP-DM from data exploration, pre-processing and transformation before applying different forecasting algorithms and evaluation. Moreover, secondary goals behind this paper include understanding associations between SKUs through market basket analysis, and clustering using KNN based on brand, customer, order quantity and value to propose a product segmentation strategy. The results of both clustering and forecasting models are then evaluated to choose the optimal forecasting technique, and a visual representation of the forecast and exploratory analysis conducted is displayed using R