Improving Sales Predictions With Autoformer By Integrating ARIMA For Anomaly Detection
Mohammad Hasan Saghafi, Monireh Moshavash, Davud Mohammadpur · 2024
Accurate sales forecasting is essential for online businesses to optimize inventory and marketing strategies. This paper addresses a significant and timely problem in e-commerce forecasting by proposing a hybrid model combining ARIMA for anomaly detection and Autoformer for long-term prediction. The integration of these models results in improved forecast accuracy, particularly in handling anomalous sales patterns. The Autoformer model, known for its performance in capturing long-term dependencies and seasonal trends, is enhanced by filtering out anomalies in the sales data using ARIMA before training. Our approach leverages historical sales data from an online shop, Technolife, to demonstrate how the detection of anomalous data leads to more accurate predictions. We provide key performance metrics such as Mean Absolute Error (MAE) and R-squared (R2) and demonstrate that handling anomalies improves forecasting accuracy.