Sales Forecasting Models: Comparison between ARIMA, LSTM and Prophet

Abstract

Sales forecasting is crucial for business planning and resource allocation. Data-driven approaches have become popular in this field. This study compares the performance of three forecasting models: Autoregressive Integrated Moving Average (ARIMA), Long Short-Term Memory (LSTM) and Prophet within the context of specific sales categories derived from acquiring data provided by a bank. This study uses a time series dataset provided by Tink off data, which covers various sales categories and time intervals. These categories, including pharmacies, railway tickets, books, sporting goods and fuel stations, present unique forecasting challenges because of their distinct demand patterns and potential for high volatility. Through a comparative analysis focusing on accuracy, robustness and computational efficiency, the study reveals that while all models demonstrate efficacy in certain scenarios, their performance varies depending on the specific category and forecasting horizon. ARIMA exhibits consistent accuracy across categories, particularly for daily predictions, aligning with its strength in capturing trends and seasonality. LSTM, on the other hand, shows promise for hourly predictions in categories like fuel stations, leveraging its ability to learn long-term dependencies. However, the LSTM model shows inconsistent results, sometimes outperforming others, but with varying performance across runs. This study provides insights for practitioners within the banking and financial sectors seeking to select the most appropriate forecasting model based on their specific sales categories and forecasting needs