Stationarity in Prophet Model Forecast: Performance Evaluation Approach

Abstract

Stationarity plays a crucial role in time series analysis, significantly influencing model performance and the reliability of forecasts. Despite its importance, many real-world datasets exhibit non-stationary behaviour, which can lead to misleading or spurious forecasting outcomes. This study explores the impact of stationarity on the performance of the Prophet Model, a scalable time series forecasting tool developed by Facebook, by comparing forecasts from both stationary and non-stationary versions of the same dataset. The monthly international airline passenger data was downloaded from the Kaggle website. We applied the Prophet to generate forecasts from raw (non-stationary) data and its transformed (stationary) version, obtained through first differencing. Three metrics, including the Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and coverage, were used to evaluate both versions of the forecasts. The findings reveal that forecasted values from stationary and non-stationary data exhibit strong correlations with actual values, as confirmed by T-tests and Pearson correlation coefficients. However, the Prophet model demonstrated notably better performance on stationary data compared to the non-stationary version, with the forecast for stationary data showing lower RMSE and MAE values and higher coverage percentages. The study shows the importance of ensuring stationarity before forecasting, even when using advanced models like the Prophet Model. We suggest that integrating stationarity considerations into future iterations of the Prophet algorithm could further enhance its predictive capabilities.