With the advent of big data and advanced analytics, management accounting practices are increasingly interested in integrating sophisticated forecasting models to enhance decision-making processes. This study embarks on a comparative analysis to assess the appropriateness and performance of three modeling methodologies—traditional time series models, machine learning driver-based models, and the Facebook Prophet algorithm—in the context of sales forecasting for management accounting purposes. My research utilizes a mixed-methods approach, combining detailed quantitative and qualitative analyses of sales forecasting outcomes from multiple retail case studies.
The findings reveal that the Facebook Prophet algorithm excels in direct (B2C) sales scenarios characterized by high data granularity and frequent data collection, offering superior accuracy and flexibility in adjusting for seasonal variations and market trends. Conversely, driver-based models demonstrate robust performance in indirect (B2B) sales scenarios where the available data does not directly reflect actual sales outcomes, such as situations with significant time lags or when data is aggregated excessively due to bulk purchases or sales. Traditional time series models serve as a useful baseline but generally underperform compared to the other methods.
This study contributes to the management accounting literature by delineating the conditions under which each forecasting model has the highest efficacy, considering factors such as data availability, external influences, robustness to change, and data aggregation levels. The implications of this research extend to the design of forecasting systems, suggesting a nuanced approach to model selection that incorporates both technological capabilities and organizational dynamics.