Production Plan Forecasting on Limited Dataset

Abstract

Forecasting production plans with limited data is a significant challenge, especially for smaller firms or new products with short production histories. This thesis aims to predict actual production outcomes based on historical data and to understand the relationships between and connect planned and executed production volumes. By exploring various forecasting approaches, including ARIMA and Long Short-Term Memory (LSTM) networks, the study focuses on methods designed to perform well with smaller datasets. The research employs a hierarchical model architecture that decomposes the forecasting task into three components: production forecasting, quarterly mapping, and unfolding to a monthly plan estimate. The models are evaluated against baseline strategies, including naive predictions and basic LSTM and ARIMA models. Results show that the proposed hierarchical models outperform baseline models, capturing the general behavior of the data more effectively. The deep learning-based model excels at capturing extremes in the time series, while the regression-based model provides stable and accurate forecasts. However, the models struggle with highly erratic production plan patterns, indicating the need for further refinement. This thesis contributes to more robust and scalable production planning solutions for data-constrained environments, offering valuable insights for both academic research and practical applications.