Forecasting Swedish Household Consumption: A Comparison of ARIMA and ARIMAX Forecasting Accuracy

Sammanfattning

This thesis investigates the role of macroeconomic variables in forecasting Swedish household consumption. Forecasts are based on quarterly data obtained from Statistics Sweden spanning from 2001 to 2025. A univariate ARIMA model is compared with three ARIMAX models containing different combinations of the macroeconomic variables employment, interest rates, income and inflation as well as a naive benchmark random walk. The first ARIMAX model includes all macroeconomic variables, the second ARIMAX model includes employment and interest rates and the third ARIMAX model includes employment and income. The forecasting methods included are a training and test split forecast and a rolling window forecast. Forecast accuracy is evaluated through a 12 quarter test period based on a 88 quarter training period. Evaluation is conducted using the RMSE, MAE and MAPE metrics. The Diebold-Mariano test is then applied to determine whether the difference is statistically significant. The study reveals that the choice of evaluation method is crucial. The results of the study suggest that inclusion of macroeconomic variables may improve forecast accuracy in a fixed training and test split evaluation as the model with all variables yielded the highest forecasting performance, although no model can be shown significantly better than the random walk. When a one-step-ahead rolling window forecast is performed, all models are outperformed by the naive random walk, although this difference cannot be proved to be significant. The study concludes that no evidence can be found whether inclusion of macroeconomic variables improves the forecast accuracy.