Nowcasting Real GDP Growth: Comparison between Old and New EU Countries

ABSTRACT

We analyze the performance of a broad range of nowcasting and short-term forecasting models for a representative set of twelve old and six new member countries of the European Union (EU) that are characterized by substantial differences in aggregate output variability. In our analysis, we generate ex-post out-of-sample nowcasts and forecasts based on hard and soft indicators that come from a comparable set of identical data. We show that nowcasting works well for the new EU countries because, although that variability in their GDP growth data is larger than that of the old EU economies, the economic significance of nowcasting is on average somewhat larger.

KEYWORDS:

• Bayesian VAR
• dynamic and static principal components
• european OECD countries
• factor augmented VAR
• nowcasting
• real GDP growth
• short-term forecasting

JEL CLASSIFICATION:

• C33
• C38
• C52
• C53
• E37
• E52

Acknowledgments

We are thankful for valuable comments we received from Tomáš Adam, Joe Brada (Editor), Ali Kutan, Ashot Mkrtchyan, Armen Nurbekyan, Nerses Yeritsyan, and participants at several presentations. Kočenda acknowledges support from the Grant Agency of the Czech Republic grant No. 19-15650S. The views presented in this paper do not represent the official views of the Central Bank of Armenia. The usual disclaimer applies.

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

1. Nowcasting is the prediction of the present, the very near future and the very recent past in economics to monitor the state of the economy in real time.

2. Our motivation is further supported by findings of Benczúr and Rátfai (2014), who cover a large sample of countries prior to 2008, that includes also those analyzed by us, and show existence of strong degree of heterogeneity in macroeconomic behavior both across and within groups of emerging and developed countries. Further, they document differences in the volatility of the cyclical component of output. In general, they show that output is about twice as volatile in emerging market countries as in industrial countries.

3. Alternative short-term forecasting algorithms include the traditional autoregression (AR) model, factor-augmented autoregression (FAAR) model, unrestricted vector autoregressive (VAR) model, small-scale Bayesian VAR (BVAR), unrestricted factor augmented VAR (FAVAR), and Bayesian factor augmented VAR (BFAVAR).

4. In our analysis, we use a nowcasting algorithm proposed by Giannone, Reichlin, and Small (2008) because a large number of central banks have adopted it successfully. Further, our aim is to compare the most popular nowcasting method with other short-term forecasting algorithms, not with alternative nowcasting algorithms, such as MIDAS or mixed-frequency VAR. A comparison of various nowcasting algorithms is a topic for future research.

5. The MATLAB codes for nowcasting are available at https://www.newyorkfed.org/research/economists/giannone/pub/, which presents the computational steps of the nowcasting algorithms proposed by Giannone, Reichlin, and Small (2008) in full detail.

6. We assume that no structural relationship exists among endogenous variables and the existence of a left-hand side unity matrix (1 on the diagonal and 0 otherwise).

7. Following Blake and Mumtaz (2012), we use the Minnesota-type prior representing the belief that endogenous variables in a VAR model follow an AR(1) process or a random walk.

8. The MATLAB codes for two-step dynamic factor model is available at https://www.newyorkfed.org/research/economists/giannone/pub/, which presents the computational steps of the time domain algorithm proposed by Doz, Giannone, and Reichlin (2011) in full detail.

9. Camacho and Perez-Quiros (2010) use five hard indicators (Euro area industrial production index, excluding construction; Euro area total retail sales volume; industrial new orders index; total manufacturing orders; extra-Euro area exports) and five soft indicators (Belgium overall business indicator, Euro-zone economic sentiment indicator, Germany IFO business climate index, Euro area manufacturing purchasing managers index, Euro area services purchasing managers index).

10. Specifically, for our set of countries, the industrial production and retail trade indices are available at least four to seven weeks after the end of the current month (1–1.5 months). Similar lags are also present for imports and exports indices (3–5 weeks).

11. The countries are members of the OECD. We do not use the full OECD membership (36 countries) because of inconsistencies in data availability for the rest of the countries and because we are concentrating here on European countries.

12. We do not conduct a backcast because we do not have flash and first-revision GDP data. That is why we can generate a nowcast with revised data. Further, we do not perform one- or two-quarter-ahead forecasts because our target is to compare nowcasting and short-term forecasting algorithms for the current quarter.

13. This strategy is a good compromise among the standard in-sample and out-of-sample proportions of 50/50, 70/30, and 90/10 employed in modern machine learning modeling (see, e.g., https://machinelearningmastery.com/backtest-machine-learning-models-time-series-forecasting/).

Additional information

Funding

This work was supported by the Grantová Agentura České Republiky [19-15650S].