China Activity Proxy – Methodology

The China Activity Proxy (CAP) is our attempt to track the pace of economic growth in China without relying on the official GDP figures. We began compiling the CAP in 2009 and expanded its scope in 2020, though the framework remains the same. A detailed analysis of how the CAP improves on the GDP data, and the advantages it offers in tracking China’s economy can be found in our Focus, “Introducing our new China Activity Proxy“.

The CAP is based on a set of carefully selected indicators that capture activity across a broad spectrum of the economy. While much of the data is collected by government entities, we have avoided using high-profile indicators such as industrial production and retail sales, focusing instead on low-level indicators that are less politically-sensitive and so less likely to be manipulated. The CAP is based solely on monthly indicators, giving a more regular and timely view of economic conditions than is possible using quarterly GDP.

THE CAP’S COMPONENTS

The CAP is made up of eight indicators:

  1. CE Industrial Output Index (units vary). Our in-house proxy for industrial production that aggregates data on the output volumes of key products (metals, chemicals, electronics, etc).
  2. Freight Traffic (tons-km). A broad measure that captures the movement of goods across China.
  3. Seaport Container Traffic (TEU). A proxy for the volume of foreign and intercoastal trade in manufactured goods.
  4. CE Construction Machinery Sales Index (units). A proxy for construction activity covering both property and infrastructure that aggregates data on the sale of machinery such as excavators, cranes, road rollers and forklifts.
  5. Passenger Traffic (persons-km). Captures leisure and business travel.
  6. Property Sales (sqm). A proxy for real estate services activity.
  7. Car Sales (units). A proxy for discretionary consumer spending.
  8. Service Sector Electricity Consumption (kwh). A proxy for broader service sector activity.

All the CAP components are measured in volume terms, allowing us to avoid problems translating nominal into real values. They extend back to at least 2007 and most go back further. We have made some adjustments to the series to accommodate year-end revisions and shifts in the timing of Lunar New Year.

THE CAP’S CONSTRUCTION

Estimation of annual (% y/y) CAP growth follows a three-stage process.

  1. First, we calculate the annual growth rates for each of the component series.
  2. Next, we calculate weights for the component series to ensure that changes in less volatile series have the same impact on the CAP as proportionally-equal changes in more volatile series. We do this by de-trending and normalising their growth rates.
  3. We then average the original growth rates using these weights. This allows us to adjust for volatility in the component series but still preserve the trends in the underlying data.

At no point in the process are the official GDP data used, whether as inputs or to weight or scale the CAP indicators.

We also construct a seasonally-adjusted version of the CAP. We use thee monthly industrial production data for 1999 as a base and then extrapolate forward using the CAP growth rates. We then seasonally adjust the resulting series using X-12-ARIMA with custom holiday regressors for Chinese New Year.

WHY WE CHOSE THIS APPROACH

Our approach of averaging normalized growth rates of component series is simple but has an important advantage over more sophisticated econometric techniques such as principal components analysis (PCA) or dynamic factor models (DFM).

PCA and DFM using the same underlying series result in an output with broadly the same trajectory of growth as our approach. But their output needs to be scaled in some way to generate an estimate for the pace of economic growth, as these approaches discard the information contained in the level of growth of the component series.

This scaling requirement is not a drawback if the goal is to “nowcast” official GDP. But our goal is to estimate actual growth, which by assumption we cannot observe. Any scaling would be arbitrary and introduce biases.