Abstract
In the fast-moving consumer packaged goods (FMCPG) industry, long-range market share forecasting provides insight into guide strategic planning and optimizes resource allocation. In collaboration with an international FMCPG company, this study evaluates multiple forecasting models to predict market share trends for their monthly U.S. cookies and crackers category over a 24-month horizon. The approach examines a range of forecasting techniques, including moving average, seasonal naïve, autoregressive integrated moving average (ARIMA, SARIMAX), linear regression, random forest, XGBoost, and hybrid boosted models. Where relevant, the models incorporate a comprehensive set of independent variables, including historical sales of competitor brands, macroeconomic indicators, competitive pricing trends, and promotional activity. Each model is evaluated through cross-validation and tested against historical data using mean absolute percentage error (MAPE) and monthly average bias as performance metrics. Results show that while machine learning models can capture nonlinear relationships, without sufficient training data they are prone to overfitting, thus limiting their use in data-limited scenarios. In contrast, time series models provide a balance of accuracy and interpretability. Using linear regression, we achieved a MAPE of 1.28% and a bias of 0.33% over a 24-month horizon. These findings demonstrate the effectiveness of structured statistical modeling in market share forecasting and provide the FMCPG company with a tool to assist in strategic planning.
