Product level analysis of Japan's trade elasticities

As a part of MACS30123 Large-Scale Computing for the Social Sciences, this personal project works on Japanese Trade Statistics and conducts Big Data analysis with Data Visualization and Regression.

Social science research problem

Determining the factors that influence trade flows has been a prominent area of study in international financial literature (refer to [1] for an overview). One specific aspect that has garnered significant attention is the exchange rate elasticity of trade balance. Understanding this elasticity is crucial as countries often focus on currency devaluation as a means to improve their trade balance. However, empirical studies examining the J-curve effect hypothesis, which explores the trade response to devaluation in both the short and long term, have produced mixed results (refer to [2] for a review of the J-curve effect).

One of the reasons for these inconsistent findings is the reliance on aggregate export or import data in many studies. To address this issue, Bahmani-Oskooee has conducted several studies using country-specific trade data or product-level data (e.g., [3,4]). This approach allows for a more nuanced understanding of the heterogeneous effects of exchange rates.

Furthermore, the integration of the global value chain (GVC) in recent years has significantly altered the trade flow response to exchange rates. For instance, Japanese manufacturers began relocating their operations abroad during the 1990s, reducing their vulnerability to exchange rate fluctuations [5].

Building upon Bahmani-Oskooee's methodology, the objective of our study is to analyze the exchange rate elasticities at the product level using up-to-date data and detailed bicountry and product level Japanese data.

Justification of the importance of using scalable computing methods

Obtaining access to detailed and granular trade statistics data from the Japanese government presents a valuable opportunity for our research. The availability of monthly data at such a high level of granularity, encompassing all trading partners and product codes for imports and exports, provides us with a comprehensive dataset. However, it is important to note that this level of granularity also results in a large dataset, with a total of 51,291,956 records.

While it may be tempting to rely on more aggregated data to simplify the analysis, doing so would lead to a loss of crucial information and subsequently reduce the precision of our estimates. To ensure that we extract the maximum insights and preserve the accuracy of our findings, it is imperative to employ large-scale computing methods throughout our research process.

Dataset

• Trade Statistics

  Update Kaggle dataset with recent data. Monthly level data from 1988-2020 and most granular product level (HS9 codes).

  • Kaggle 1988-2020
  • E-Stat: scarped with API

• Exchange Rates

  Downloaded monthly data of "National Currency Per U.S. Dollar, Period Average Rate" from IMF. Eventually converted to National Currency Per Japan Yen in our analysis.

• GDP

  Downloaded quarterly data of "Nominal, Domestic Currency, Seasonally Adjusted GDP" from IMF.

Since their is a mismatch between country codes between trade dataset and IMF dataset, we used `fuzzywuzzy' to match country names and then manually modify some mismatch.

Model

We run Panel ARDL model (Autoregressive distributed lag model) with linearmodels package. Particulary, Panel OLS is used based on [3] and [4], which can be described as following.

$$ \begin{align} \Delta\log(Y_{i,c,t}) &= \delta_0 + \lambda_0\log(Y_{c,t-1})+\lambda_1\log(EX_{c,t-1})+\lambda_2\log(GDP_{c, t-1})\\ &+\sum_{j=1}^3\phi_j\Delta\log(Y_{t-j})+\sum_{j=0}^3\rho_j\Delta\log(EX_{t-j})+\sum_{j=0}^3\eta_j\Delta\log(GDP_{t-j})\\ &+\alpha country_c+\beta time_t+ \varepsilon_{c,t} \end{align} $$

The subscript $i$ is for product, $c$ is for country, and $t$ is for time. $Y$ is the value of trade flows of export and import respectively. $EX$ is exchange rate (Domestic Currency/Japan Yen). $GDP$ in the first equation is trading's partner GDP. $country$ is for entity fixed effect and $time$ is for time fixed effect. Our dataset contains 88 unique hs2 codes so that we run 176 regressions. The time period is 1988Q1-2023Q1. Missing value is droped.

We are interested in exchange rate elasticities. We can derive short and long-run coefficients following [3] and [4].

$$ \begin{align} \text{short-run coefficients}&: \frac{\rho_j}{\phi_j}\quad(j=1,2,3)\\ \text{long-run coefficient}&: \frac{\lambda_1}{\lambda_0} \end{align} $$

Pipelines

• Scraping of recent trade statistics: scraping.ipynb

  • This is done in local and not parallelize due to API's limitation.

• Upload data to S3: upload_data.ipynb

  • Upload raw data files to S3. Some manipulation is done for small data (IMF's exchange rate and GDP data).

• Processing trade data with pyspark: data_manipulation_pyspark.ipynb

  • Process old data from Kaggle and new data we scraped so that can be merged later.
  • We chose PySpark because it was impossible to handle large data with memory issue in our local environment. We also tried Dask but PySpark was a lot faster to do basic data manipulation. We configure PySpark environment as follows.

  aws emr create-cluster \
  --name "Spark Cluster" \
  --release-label "emr-6.2.0" \
  --applications Name=Hadoop Name=Hive Name=JupyterEnterpriseGateway Name=JupyterHub Name=Livy Name=Pig Name=Spark Name=Tez \
  --instance-type m5.xlarge \
  --instance-count 6 \
  --use-default-roles \
  --region us-east-1 \
  --ec2-attributes '{"KeyName": "vockey"}' \
  --configurations '[{"Classification": "jupyter-s3-conf", "Properties": {"s3.persistence.enabled": "true", "s3.persistence.bucket": "trade-final-project-bucket"}}]'
  

• EDA with pyspark: EDA_pyspark.ipynb

  • Merge old data from Kaggle and new data we scraped.
  • Add country and HS code information.
  • Data Visualization.
  • Similarly used Pyspark to work on large dataset.

• Prepare data for ARDL analysis with pyspark: data_manipulation_ARDL_pyspark.ipynb

  • Merge trade data with exchange rate data (monthly data).
  • Aggregate monthly data to quarterly data and hs9 codes to hs2 codes. Sum export value and average exchange rate.
  • Merge quarterly data with GDP data.
  • Change numerical data to log base.
  • Add lag and difference variables.
  • Similarly used Pyspark to work on large dataset.

• Run ARDL with Dask: ARDL_dask.ipynb

  • Run multiple ARDL regression for import and export for all hs2 codes.
  • Extract model summary and calculate coefficients.
  • Use Dask because ARDL model is available in latest statsmodels library and we could not install to Pysark. We configure Dask environment as follows where bootstrap code is stored here.

  aws emr create-cluster --name "Dask-Cluster" \
   --release-label emr-6.2.0 \
   --applications Name=Hadoop \
   --instance-type m5.xlarge \
   --instance-count 6 \
   --bootstrap-actions Path=s3://trade-final-project-bucket/bootstrap \
   --use-default-roles \
   --region us-east-1 \
   --ec2-attributes '{"KeyName":"vockey"}'
  

Results

EDA

First, we conduct some basic data visualization to check the data is properly collected.

1. Top 10 trading partners

   • The U.S. and China are the major trading partner in both export and import.
   • Japan exports to many asian countries (China, Korea, Taiwan, Hong Kong, Thailand, Singapore, and Malaysia), while it imports from resource-rich countries (United States, Australia, Saudi Arabia, and United Arab Emirates)

2. Time series of trade flows value in Top 10 trading partners

   • The trades with China have grown rapidly.
   • The U.S. has been a stable trading partner over the year.
   • Large drops in trade value are observed around 2009 and 2020 due to crises.

3. Time series of top 10 products in HScode basis.

   • HS2 code 84, 85 and 87 (see legend for detail) show similar trend in export value.
   • Mineral fuel is the major imported product. However, the trade flow is volatile.
   • Some HS2 codes are common between export and import (29, 84, 85, 87, and 90). Probably this is because of GVC.

_{Upper graph
84:Machines, 90: Optical photographic cinematographic measuring checking precision medical or surgicalinstruments and apparatus; parts and accessories thereof, 29: Organic chemicals, 39: Plastics and articles thereof, 19: Preparations of cereals flour starch or milk; pastrycooks'products, 85: Electrical machinery and equipment and parts thereof; sound recorders and reproducers television image and sound recorders and reproducers and parts and accessories of such articles, 87: Aircraft spacecraft and parts thereof, 72: Iron and steel, 73: Articles of iron or steel, 89: Ships boats and floating structures}

_{Lower graph
27: Mineral fuels mineral oils and products of their distillation; bituminous substances; mineral waxes, 90: Optical photographic cinematographic measuring checking precision medical or surgical instruments and apparatus; parts and accessories thereof, 29: Organic chemicals, 84: Machines, 44: Wood and articles of wood; wood charcoal, 26: Ores slag and ash, 87: Vehicles other than railway or tramway rolling-stock and parts and accessories thereof, 71: Natural or cultured pearls precious or semi-precious stones precious metals metals clad with precious metal and articles thereof; imitation jewellery; coin, 85: Electrical machinery and equipment and parts thereof; sound recorders and reproducers television image and sound recorders and reproducers and parts and accessories of such articles, 30: Pharmaceutical products.}

ARDL results

The estimation results of the ARDL models are presented below. The export table is sorted in ascending order based on the long-term coefficient. In theory, if Japan's currency depreciates, exports should increase in the long term. In our model, the exchange rate is defined as the trading partner's currency divided by the Japanese Yen. The coefficients should be negative if the hypothesis is true. However, the number of products with negative coefficients is lower than those with positive coefficients. Among the major products highlighted in the data visualization, product codes 72, 84, and 90 have positive coefficients, while the remaining coefficients are insignificant. Additionally, we observe that the absolute values of the coefficients are smaller than those reported in [3] for the United States. This observation is consistent with [5], which suggests a decreasing significance of trade elasticity in Japan.

Regarding the import results, in contrast to the export case, the positive coefficient values align with the theoretical expectations. However, out of the 87 products, only 19 have coefficients that are statistically significant, similar to the findings in [3]. It is worth noting that among the major HS2 codes, only product codes 30 and 44 have coefficients with significant values, but their signs are opposite.

Please note that there are several limitations in this estimation. Firstly, in the import model, it would be more appropriate to include Japan's GDP instead of the trading partner's GDP. Secondly, we present the p-values for $\lambda_1$ and $\rho_j$ instead of normalized values. This is due to the limited flexibility of the Python package used for analysis, and it is better to use Stata for conducting this analysis.

Export Results

Import Results

Disscussion and Conclusion

Our project findings suggest that the relationship between trade flows and exchange rates is weaker than expected, providing limited support for the J-curve hypothesis. This insight is crucial for policymakers as it indicates that there are other significant factors influencing trade flows. While we did not delve into specific reasons within the scope of this study, industry structures and demand might play a crucial role in determining trade flows.

From a large-scale computing perspective, employing Pyspark proved to be highly effective in handling large datasets. However, working with econometric models posed challenges as neither PySpark nor Dask support the majority of models available in Stata or R. Therefore, considering the utilization of Stata in the Midway2 environment might be a more suitable option for future endeavors involving econometrics.

In conclusion, our project contributes to the understanding of the intricate relationship between exchange rates and trade flows with granular data. While the J-curve hypothesis did not hold strong empirical support, it highlights the importance of considering other factors influencing trade dynamics. Further research and analysis are necessary to explore and identify the structural reasons within industries and demand that exert significant influence on trade flows.

References

[1] Krugman, P., Obstfeld, M., Melitz, M. 2018. International Economics: Theory and Policy. Pearson Education Limited.

[2] Bahmani-Oskooee, M., and Ratha, A. 2004. The J-curve：a literature review. Applied economics, 36(13), 1377–1398.

[3] Bahmani-Oskooee, Mohsen, and Zohre Ardalani. 2006. Exchange Rate Sensitivity of U.S. Trade Flows: Evidence from Industry Data. Southern Economic Journal 72, no. 3 (2006): 542–59.

[4] Bahmani-Oskooee, M., & Aftab, M. 2017. On the asymmetric effects of exchange rate volatility on trade flows: New evidence from US-Malaysia trade at the industry level. Economic Modelling, 63, 86-103.

[5] Shimizu, J., Sato, K. 2015. Abenomics, Yen Depreciation, Trade Deficit, and Export Competitiveness. RIETI Discussion Paper,15-E-020.