There are several provinces with zero search volume indices (SVI) during the first years of the data range examined in this study, and this is a limitation of the Google Trends data due to lower search intensity. This situation will affect the performance of the transmission now, so it is necessary to evaluate the benefit of the current model in that period. One of the most obvious sources of big data provided by Google is the Google Trends data.
The first component of principal component analysis is used to group subsectors into GRDP sectors. The factor loading matrix explaining the contribution of the variable to the common factor is expressed as Λ. In this context, unbounded means that the polynomial lag function of the high-frequency variable is not included.
To prevent the spread of the virus, the government imposed social restrictions in various provinces at the beginning of the pandemic. This means that none of the variables in West Kalimantan has a correlation value exceeding 0.23. At the national level, we use variables that are commonly used at the national level for studies now: export, import, narrow money, broad money, manufacturing industry index, general wholesale price index, and foreign tourists.
However, the results at the national level are smoother and more accurate due to various data problems at the regional level.
North Sumatera
This figure shows that the nowcasting model can cushion the economic downturn of the COVID-19 pandemic. In the first quarter of 2020, a low magnitude, along with the economic shock, was demonstrated by the nowcasting results. However, in the fourth quarter of 2020, economic movements in the seven provinces started to expand and the nowcasting models also showed a revival of the economic pattern.
Regional level variables are the sets of variables that we will use in our regional level study. M1 is narrow money, M2 is wide money, IBS is index of manufacturing industry, IHPB is index of general wholesale prices, foreign tourists are all foreign visitors who arrived directly in Indonesia. FIN refers to financial variables, MACRO represents macroeconomic variables, and GT represents Google Trend variables.
This figure shows the year-over-year growth rate of GRDP (black line) compared to DFM nowcast using financial and macroeconomic variables (blue line) and Google Trends variables (red line) in 7 Indonesian provinces.
Lampung
Central JavaD. Banten
Bali
West Sumatra
Jambi
Bangka Belitung Islands
South Kalimantan
The results of GRDP forecast in the North Sumatra, West Sumatra, Jambi, Riau Islands, Central Java, Banten, Bali, East Java and South Kalimantan indicate that adding more observations for GRDP forecast in 2019Q4 did not significantly increase the forecast accuracy between the forecast periods. This condition also occurs in several provinces experiencing economic shocks, namely Riau Islands and South Kalimantan. DFM RMSE of Nowcasting using financial and macro sets of variables by methodologies and provinces.
The note indicates the number of months before the quarterly GDP value is issued. Q4 2019 refers to the current moment of the fourth quarter of 2019, the period before the COVID-19 pandemic. One of the reasons behind this is that this model does not differentiate between handling non-linearity in the data.
Woloszko's (2020) revealed similar findings for a period including the economic shock due to the global economic crisis in 2008, showing that forecasting results have worsened. In the sample period up to the fourth quarter of 2019, which excludes the economic shock due to the COVID-19 pandemic, forecasting with Google Trend variables produces better forecasts in all provinces except in the case of Riau, Lampung and East Java than those financial and macroeconomic variables. . When compared with the univariate model, our results suggest that North Sumatra, Banten, and Bali outperformed the AR model, while Riau, Lampung, Central Java, and East Java have higher accuracy using the AR model.
This means that the Google Trend variables have been able to provide meaningful information in cases where the RMSE of those provinces is only slightly different from the others, except in Riau. This shows that, in the presence of stable economic conditions, Google Trend variables can be effectively used to explain the direction of economic movement. Overall, the current MIDAS models based on the financial, macroeconomic, Google Trend variables shown in Figure 4 show the direction of economic decline since the pandemic.
Some multi-province models may well depict the decline in economic growth, which reached negative values in all provinces in the second quarter of 2020. Moreover, economic recovery movements can be captured by nowcasting in the fourth quarter of 2020, when the expansion begins. This figure shows the year-on-year growth rate of GRDP (black line) compared to MIDAS nowcast using financial variables (blue line), macroeconomic (green line) and google trend variables (red line) in 7 Indonesian provinces.
North Sumatra
Riau
Banten
East Java
Overall, the nowcasting results in Figure 5 show that the MIDAS model accurately captures the decline in economic activity that occurred in those six provinces during the COVID-19 pandemic. Except for West Sumatra, almost all models with macroeconomic variables report negative magnitudes in the second quarter of 2020. Several regions, such as West Sumatra and the Bangka Belitung Islands, witnessed an increase towards the nowcasting in the fourth quarter of 2020, as Jambi continued to fall.
Riau Islands
West Kalimantan
CONCLUDING REMARKS
Based on the narrowcast results using a selected set of variables, this study concludes that narrowly distributed quarterly GRBP growth at the provincial level in Indonesia can be achieved using macroeconomics, financial and Google Trends data. Since the COVID-19 pandemic started in the first quarter of 2020, the DFM and MIDAS models are able to capture the decline in economic activity. The results of nowcasting using the DFM method have been adjusted to economic growth in the fourth quarter of 2020.
Apart from the macroeconomic variables of Jambi and Bangka Belitung islands, the MIDAS model also produces negative magnitudes. The DFM and MIDAS model performance shows better accuracy in a data range that does not include periods of economic shocks. Model accuracy across the same set of variables and provinces is slightly reduced when the pandemic period is included.
Additionally, using the Google Trends variable provides the best accuracy compared to other categories of variables using the MIDAS model, both before and after the COVID-19 pandemic. Using the DFM model, the use of Google Trends variables also gives the best accuracy, but only in the pre-pandemic period. The process of forecasting current times can be adopted by researchers and policy makers, including the central bank and local government.
Therefore, the local government can adapt the method of the Federal Reserve Bank of New York, which publishes the latest current results on their official website. This can be useful for provincial development plans as well as information for policy makers. Forecasting short-run real GDP growth in the Eurozone and Japan using unrestricted MIDAS regressions.
