Forecasting Portugal's GDP with ARIMA
Box-Jenkins time series model of Portugal's quarterly GDP, compared against EC, IMF and OECD forecasts. NOVA IMS course project, 3-person team.
ARIMA(3,1,0) won on the hold-out period with a MAPE of about 5.15%.
Problem
How far can a model that only sees past GDP get compared with the forecasts of large institutions?
Approach
Eurostat quarterly GDP, 1995 Q1 – 2025 Q4, seasonally adjusted. Trained up to 2022 Q4 and held out 2023 Q1 – 2025 Q4. Log-differenced the series and checked stationarity with the ADF test. Selected candidates by AIC and BIC, and checked residuals with ACF plots and the Ljung-Box test.
Result
ARIMA(3,1,0) beat ARIMA(0,1,3) on the test period. Refitted on all data, it forecasts nominal growth of 5.47% (2026), 4.46% (2027) and 4.33% (2028). That is slightly above the institutions for 2026, with the same slowdown in 2027.