Optimizing Sales Forecasting in e-Commerce with ARIMA and LSTM Models

Sales forecasting is the process of estimating future revenue by predicting the amount of product or services a sales unit will sell in the near future. Although significant advances have been made in developing sales forecasting techniques over the past decades, the problem is so diverse and multi-dimensional that only in a few cases high accuracy predictions can be achieved. In this work, we propose a new hybrid model that is suitable for modeling linear and non-linear sales trends by combining an ARIMA (autoregressive integrated moving average) model with an LSTM (Long short-term memory) neural network. The primary focus of our work is predicting e-commerce sales, so we incorporated in our solution the value of the final sale, as it greatly affects sales in highly competitive and price-sensitive environments like e-commerce. We compare the proposed solution against three competitive solutions using a dataset coming from a real-life e-commerce store, and we show that our solution outperforms all three competing models.