An Investigation into the Performance of Time Series Models in Predicting US E-commerce Data

In recent years, the e-commerce data landscape in the US has undergone significant transformations, especially under the shock of the COVID-19 pandemic, reflecting broader shifts in consumer behaviour and technological advancements as the digital marketplace continues to expand. Consequently, efficiently predicting e-commerce retail sales data becomes increasingly crucial. This paper evaluates the effectiveness of time series models in forecasting the non-seasonally adjusted US e-commerce retail sales data, thereby comparing the ability of these models to capture the characteristics of the US e-commerce market, such as seasonality which is an important factor for e-commerce yet had not been sufficiently researched. This paper covers time series models including the Autoregressive Integrated Moving Average (ARIMA) model, the Error, Trend, and Seasonality (ETS) Additive and ETS Multiplicative models. Two different splitting methods of training and testing sets are implemented on the data to analyse the impact of the pandemic on the performance of these models. The ARIMA model is the best-performing model under both splitting methods as it produces the best Root Mean Square Error (RMSE) and residuals. However, its prediction accuracy is much lower under the splitting method where the testing window is affected by the COVID-19 pandemic. Government restrictions, consumer behaviour shifts and the financial fragility of businesses are likely to be the factors contributing to the sudden shift in the e-commerce retail sales data.