As stated by Dauphin et al (2022), The basic idea of nowcasting is to exploit a diverse set of timely information available before the official release of a target variable. Therefore, data selection and transformation are key to the success of nowcasting. In this research, the data used for the estimation consists of monthly indicator data from both macroeconomic and financial perspectives, with an observation period spanning from 2013 to 2023. There are a total of 34 variables, including retail sales value and tax income, with more of these variables listed in Table 1. Subsequently, the various data will be transformed into year-on-year percentages to prevent spurious relationships. It is important to note that the data used within a specific time range will have a one-month lag. For example, nowcasting Q4 2023 GDP growth in January 2024 will use October, November and December 2023 data. While the official data for GDP Q4 itself will be released on the first week of February.
The projection evaluation will be divided into two parts: in-sample and out-of-sample projections. The in-sample projections use training data from the first month of 2013 (2013M01) to Desember 2022 (2022M12), while the out-of-sample projections will use testing datasets from January (2023M01) to Sepember 2023 (2023M09). For future evaluations of the model, we will periodically use new data released by the Central Bureau of Statistics to asess the accuracy of the projection result.
Table 1
Code
|
Variable
|
Frequency
|
Data Lag
|
Unit
|
gdprl
|
GDP of Indonesia
|
Quarterly converted to Monthly by Constant Method
|
1 Month
|
YoY Percentage (%) Change
|
retailsales
|
Retail Sales
|
Monthly
|
1 Month
|
YoY Percentage (%) Change
|
vehicleparts
|
Vehicle Parts
|
Monthly
|
1 Month
|
YoY Percentage (%) Change
|
foodbeverages
|
Food and Beverages
|
Monthly
|
1 Month
|
YoY Percentage (%) Change
|
autofuels
|
Fuels of Automotive
|
Monthly
|
1 Month
|
YoY Percentage (%) Change
|
infocomequip
|
Information and Communication equipment
|
Monthly
|
1 Month
|
YoY Percentage (%) Change
|
hholdequip
|
Household Equipment Index
|
Monthly
|
1 Month
|
YoY Percentage (%) Change
|
recreationgood
|
Recreation Goods
|
Monthly
|
1 Month
|
YoY Percentage (%) Change
|
othergoods
|
Other Goods
|
Monthly
|
1 Month
|
YoY Percentage (%) Change
|
clothinggoods
|
Clothing Goods
|
Monthly
|
1 Month
|
YoY Percentage (%) Change
|
mobilsales
|
Automobile Sales
|
Monthly
|
2 Month
|
YoY Percentage (%) Change
|
motorsales
|
Motorcycle Sales
|
Monthly
|
2 Month
|
YoY Percentage (%) Change
|
prod_motor
|
Motorcycle Production
|
Monthly
|
2 Month
|
YoY Percentage (%) Change
|
pmi
|
Purchasing Managers Index
|
Monthly
|
1 Month
|
YoY Percentage (%) Change
|
farmertradeidx
|
Index of Farmer Trade
|
Monthly
|
1 Month
|
YoY Percentage (%) Change
|
idx_ihsg
|
Composite Stock Price Index
|
Monthly
|
1 Month
|
YoY Percentage (%) Change
|
consconfidx
|
Consumer Confidence Index
|
Monthly
|
1 Month
|
YoY Percentage (%) Change
|
curreconidx
|
Current Economy Index
|
Monthly
|
1 Month
|
YoY Percentage (%) Change
|
consexpctidx
|
Consumer Expectation Index
|
Monthly
|
1 Month
|
YoY Percentage (%) Change
|
currincomeidx
|
Current Income Index
|
Monthly
|
1 Month
|
YoY Percentage (%) Change
|
jobavailidx
|
Job Availibility Index
|
Monthly
|
1 Month
|
YoY Percentage (%) Change
|
purchdurableidx
|
Purchase of Durable Goods Index
|
Monthly
|
1 Month
|
YoY Percentage (%) Change
|
rtgstx
|
Real Time Gross Settlement (RTGS) Transaction Value
|
Monthly
|
1 Month
|
YoY Percentage (%) Change
|
skntx
|
National Clearing System Transaction Value
|
Monthly
|
1 Month
|
YoY Percentage (%) Change
|
marketcap
|
Market Capital
|
Monthly
|
1 Month
|
YoY Percentage (%) Change
|
idx_lq45
|
LQ45 Stock Index
|
Monthly
|
1 Month
|
YoY Percentage (%) Change
|
idx_basic_ind
|
Basic Industry Stock Index
|
Monthly
|
1 Month
|
YoY Percentage (%) Change
|
idx_infr
|
Infrastructure Index
|
Monthly
|
1 Month
|
YoY Percentage (%) Change
|
idx_finance
|
Financial Index
|
Monthly
|
1 Month
|
YoY Percentage (%) Change
|
reserve
|
Reserve
|
Monthly
|
1 Month
|
YoY Percentage (%) Change
|
exrpl
|
Exchange Rate of Indonesia
|
Monthly
|
1 Month
|
YoY Percentage (%) Change
|
crude_oil
|
Crude Oil
|
Monthly
|
1 Month
|
YoY Percentage (%) Change
|
nontaxincome
|
Income from Non-Tax
|
Monthly
|
1 Month
|
YoY Percentage (%) Change
|
taxincome
|
Income from Tax
|
Monthly
|
1 Month
|
YoY Percentage (%) Change
|
l1prod_motor
|
Lag 1 month Motorcycle Production Volume
|
Monthly
|
1 Month
|
YoY Percentage (%) Change
|
Source: Internal |
The historical GDP data (Fig. 1) provides essential input for the entire projection. According to the Central Bureau of Statistics, Indonesia’s GDP growth rate remained relatively stable in the range of 5.0–6.0% (year-on-year percentage change) from 2013 until the first appearance of Covid-19 cases. However, from 2020 until 2021Q1, GDP growth dropped to the range of -5.0% to -0.6%. Subsequently, Indonesia GDP returned to its pre-pandemic levels in the third quarter of 2022.
The main ideas behind techniques applied in this study include: (i) determining the variables that have most strong linkages or relationship to GDP, (ii) estimating GDP values using the combined variables selected based on their relationship strength, (iii) asessing the accuracy of projection result through the in-sample and out-sample data, (iv) nowcasting the GDP using the selected variables, and (v) calculating the ensemble average from the results of all models. In depth, 34 high frequency indicators were exercised on 6 types of correlation (MIC, Max NMI, Kendall Correlation, Spearman Correlation, Pearson Correlation, and Distance Correlation). Subsequently, variables that are closely related to GDP were selected using a clustering method called normal mixture modeling. The best combination of these variables will then be estimated iteratively using four machine learning models (Elastic Net, Random Forest, XGBoost, and Support Vector Machine) and complemented by ensemble model results to obtain predictions of GDP growth in the current quarter. Figure 2 describes the methodology applied using a flow chart.
Furthermore, the Root Mean Squared Error (RMSE) is used to measure the error given by projected in-sample and out-sample data for each method to determine the best GDP Nowcasting Model. Consequently, future evaluations of the model are regularly executed to asess the accuracy of the projection result using the latest realization data. Details on association/relationship measure, clustering methods, model selection and ensamble models will be explained further on Appendices.