Home Appliance Demand Forecasting: A Comparative Approach Using Traditional and Machine Learning Algorithms

The manufacturing industry is considered one of Ecuador's most important productive sectors because it is an excellent employment and national income source. Durable consumer goods such as white and brown goods have shown an increase with a positive trend on their GDP, so there is an expectation of growth in the market in the following years. The profitability of this industry depends on various internal factors, such as supply chain management, and external factors, such as market dynamics, which subsequently allow for generating demand forecasts. This scientific article uses sales data from an Ecuadorian white goods manufacturer company to forecast demand in two production lines. The KDD methodology was used for data processing and model construction. Three classic forecasting methods were used in the experimentation: Simple Moving Average, Simple Exponential Smoothing, and ARIMA, and three forecasting methods that use artificial intelligence algorithms: Random Forest, K-Nearest Neighbors, and Artificial Neural Networks. The performance of the forecast models was evaluated using four error metrics: MSE, MAE, RMSE, and MASE. The first experiment considered all the observations in the dataset, while for the second experiment, the dataset was partitioned into training and test sections for cross-validation. Based on the results of error metrics, ARIMA is the best-performing model for the classic algorithms and Random Forest for the Machine Learning models. Machine Learning models generally show a superior performance of up to 30% compared to classical forecasting methods to generate demand forecasts for household appliances.