Nowcasting Indonesia

GDP for 2013 and 2014 based on the 2000 base year are preliminary figures projected by the statistical office: https://​www.​bps.​go.​id/​linkTabelStatis/​view/​id/​1217.

A detailed explanation of the System of National Accounts (SNA) can be found at http://​unstats.​un.​org/​unsd/​nationalaccount/​sna.​asp.

In fact, even nominal GDP data for the overlapping time periods are not comparable between the different bases for GDP series. Data are available in Table VII.1 at http://​www.​bi.​go.​id/​en/​statistik/​seki/​terkini/​riil/​Contents/​Default.​aspx.

A non-exhaustive list of countries and papers is: the USA (Bańbura et al. 2011, 2013; Giannone et al. 2008), the Euro Area (Angelini et al. 2011; Bańbura and Rünstler 2011), Germany (Marcellino and Schumacher 2010), France (Barhoumi et al. 2010), Ireland (D’Agostino et al. 2012), Norway (Aastveit and Trovik 2012; Luciani and Ricci 2014), China (Giannone et al. 2014), Brazil (Bragoli et al. 2015), New Zealand (Matheson 2010), the Global Economy (Matheson 2013), and Latin America (Liu et al. 2012).

The implicit assumption here is that since based on their expectations on future fundamentals analysts allocate their investments, they know which series to monitor to form appropriate expectations on GDP growth.

As shown by De Mol et al. (2008) since there is a lot of comovement among macroeconomic data, the set of indicators selected with statistical criteria is extremely unstable.

As the index compiled by Danareksa was not available to us, we experimented with the household consumer confidence index compiled by the Bank of Indonesia. The latter, however, displays a trending pattern that appears difficult to reconcile with the state of the economy, and we hence discarded it.

For the policy rate, we adopted the assumption that it is observed the first day of the month following the reference month. For example, the policy rate for January is observed on February 1. Of course this is an approximation because we know what the policy rate is everyday in January. In principle, we could have accounted for daily observations in the interest rate since DFMs allow us to do so (Modugno 2014). However, Bańbura et al. (2013) have shown that including data at the daily frequency is not particularly useful for nowcasting GDP, so we adopted the convention that the interest rate is monthly and is observed on the first day after the reference month.

See, for example, Maćkowiak (2007), and Raghavan and Dungey (2015), for applications to a set of emerging economies, and Kasri and Kassim (2009), and Kubo (2009), for Indonesia specifically.

It is worth noting here that the use of y-o-y transformations likely add persistence also in the idiosyncratic component so that a higher order autoregressive process might be more appropriate. However, increasing the order of the autoregressive process of the idiosyncratic component from an AR(1) to an AR(2) implies adding eleven extra states to the model. In our case, where we have few time series observations, and the quality of the data is doubtfull, adding eleven extra states in the Kalman Filter is feasible, but for sure costly in terms of computation accurancy. In summary, although an AR(1) may be not enough for capturing the whole dynamic in the idiosyncratic process, the cost of adding an extra lag is way higher than the benefits in terms of forecasting accuracy.

An alternative to the algorithm proposed by Bańbura and Modugno (2014) is the one proposed by Jungbacker et al. (2011).

As pointed out by Baffigi et al. (2004), differently from DFMs, bridge models are not concerned with particular assumption underlying the DGP of the data, but rather, the inclusion of specific explanatory indicators is based on the simple statistical fact that they embody timely updated information about the target GDP growth series.

Note that the AR(2) and the RW do not update the prediction beyond month 2 of the quarter, but update it between month 1 and month 2 due to the GDP data release. Also, the fact that the RMSE of the AR(2) and the RW in “Backcast” month 1, and “Nowcast” month 2 is different is an artifact due to the fact that when we “Backcast” our target variable is GDP from Q4 2007 to Q3 2014, whereas when we “Nowcast” our target variable is GDP from Q1 2008 to Q4 2014. Of course, had the evaluation period been longer, we would not have this problem.

The average share of exports to GDP in Indonesia in 2010–2014 was about 20% and even slightly lower for imports (at 19%). Trade data, however, provide information about activity across sectors in the economy, and the results in Table 4 show that this is useful information for predicting GDP.

Note that this is all the “real-time” information that we have available since, as mentioned in Sect. 4.1, we do not have any “real-time” information for the other variables in the dataset.

Let \(X^y_q=100\times \log (\hbox {GDP}^y_q)\) be GDP of the q th quarter of year y, and let \(Z^y=100\times \log (\hbox {GDP}^y)\) be GDP of year y. Then, by definition \(x^y_q=X^y_q-X^{y-1}_{q}\) is the y-o-y growth rate, while \(z^y=Z^y-Z^{y-1}\) is the annual growth rate. Following Mariano and Murasawa (2003), we make use of the approximation \(Z^y\approx (X^y_1+X^y_2+X^y_3+X^y_4)/4\), which allow us to write the annual growth rate as a function of y-o-y growth rates: \(z^y=Z^y-Z^{y-1}\approx (X^y_1+X^y_2+X^y_3+X^y_4)/4 - (X^{y-1}_1+X^{y-1}_2+X^{y-1}_3+X^{y-1}_4)/4 = (x^y_4+x^y_3+x^y_2+x^y_1)/4\).

Consensus Economics Inc. forecasts comprise quantitative predictions of private sector forecasters. Each month survey participants are asked for their forecasts of a range of macroeconomic and financial variables for the major economies.

By using the LARS algorithm as in Bai and Ng (2008), we selected 10 out of all available indicators that were retrieved for Indonesia.

With the exception of few technical and minor details, this is essentially the same model used by Giannone et al. (2008). In practice, every time we update the prediction we use a balanced panel to estimate the factor with PCA. Then we fit an AR(2) on the estimated factor and use the Kalman Filter to account for missing values at the end of the sample. Finally we estimate Eq. (4) with OLS. An alternative way to perform this exercise would be to use the collapsed dynamic factor model of Bräuning and Koopman (2014).