Drivers of Portfolio Flows into Chinese Debt Securities Amidst China’s Bond Market Development

2.1 Related Literature

A vast body of empirical literature exists regarding drivers of capital flows to emerging markets. Calvo et al. (1993) and Fernández-Arias (1996) were the leaders in studying the contributions of external push factors and country-specific pull factors as drivers of international capital flows. The former analyzed flows in the Latin American context, while the latter focused on a sample of 13 countries (China not included). More recently, Koepke (2018b) reviewed and summarized key findings of subsequent studies that have found evidence on the significance of both factors in driving capital flows into emerging market economies. For debt flows particularly, there is strong evidence for the importance of push factors and mixed evidence for pull factors being significant drivers of flows. Yeyati and Zuñiga (2016) and Hannan (2018) also provide comprehensive reviews on the literature regarding push and pull drivers of capital flows. Nonetheless, as pointed out by Koepke (2018b), empirical results may differ across studies on the back of different country samples or definitions of which countries are classified as emerging markets. Therefore, comparing studies and drawing country-specific conclusions from the existing literature is difficult.

Push factors help determine the supply of capital seeking diversification opportunities abroad. They include such variables as interest rates in advanced economies, particularly the US, or corresponding rate expectations, international investors’ risk appetite, and the output growth of mature economies. For instance, low advanced economy interest rates trigger search-for-yield behaviour, pushing portfolio flows into emerging markets and vice versa (Fernández-Arias, 1996; Ghosh et al., 2014; Byrne and Fiess, 2016; Koepke, 2018b; Li et al., 2018). Similarly, Koepke (2018) and Dahlhaus and Vasishtha (2020) show that international investor expectations of US Federal Reserve monetary policy were a significant driver of bond flows. Elevated international investors risk appetite is another push factor that drives portfolio investments into EMs, while higher risk aversion triggers capital outflows from EMs (Baek, 2006; Milesi-Ferretti and Tille, 2011; Fratzscher, 2012; Ahmed and Zlate, 2014; Ghosh et al., 2014; Giordani et al., 2017; Mercado, 2018).

Pull factors typically reflect economic activity in emerging markets, asset return prospects, and country risk indicators. Favourable conditions in the destination country (e.g., strong economic growth, sound domestic economic fundamentals, or higher country credit ratings) imply that an emerging economy is an appealing investment destination on the back of higher potential profits, pulling capital flows into its financial markets (Ghosh and Ostry, 1993; Chuhan et al., 1998; Mercado and Park, 2011; Giordani et al., 2017). In a closely related study focused on surges of inflows to emerging markets (including China) and their determinants, Ghosh et al. (2014) find that recipient countries’ fundamental factors such as capital account openness, an exchange rate regime and external financing needs, influence the magnitude of capital flow surges. Meanwhile, Capolare et al. (2022) add to the push and pull factor literature, showing that news media coverage is an important determinant of cross-border portfolio flows, particularly for bond inflows and outflows into and from the US. Similarly, Choi et al. (2023) show that domestic uncertainty also functions as a local pull factor, with elevated uncertainty in the destination country (including China) reducing foreign capital inflows, particularly bank credit and portfolio bond inflows.

Push and pull factors are not static but tend to vary over time based on global and domestic economic conditions (Milesi-Ferretti and Tille, 2011; Fratzscher, 2012; Erduman and Kaya, 2016; and Lo Duca, 2012). Lo Duca (2012), for instance, studies a group of eight EM countries including China and shows that push and pull factors are subject to substantial time variation, with major changes in the importance of the drivers of flows coinciding with notable market events or shocks. The analysis by Fratzscher (2012) of portfolio flows to 50 advanced and emerging market economies (including China) similarly finds that push factors were more dominant than pull factors during the 2008–2009 global financial crisis, yet during the recovery period portfolio flows were driven more strongly by pull factors such as macroeconomic fundamentals, quality of institutions, and policies of recipient countries.

2.2 Theoretical Model for Push and Pull Factors

The existing capital flow literature typically leans heavily on empirical studies without a solid backing of economic theory. For example, the inclusion of interest rates and risk aversion variables in estimations is simply justified by established practise. Here, we take a more comprehensive approach, deriving a simple small open economy model to justify the push-pull framework. While model presented here is similar to the conceptual framework of Cerdeiro and Komaromi (2021), who study capital account restrictions and their impact on capital flows, it is more straightforward and focuses on justifying the inclusion of domestic economic growth, international interest rates, and risk aversion as explanatory variables in empirical estimations. As such, the model does not attempt to fully capture real world financial flows; rather, it functions as a simplified theoretical foundation and a starting point for the subsequent empirical analysis on the drivers of Chinese bond flows.

3.1 Empirical Model and Data

We follow established practise in the push-pull literature (Koepke 2018) and estimate variants of the subsequent general model for China’s debt portfolio flows:

where Flows_t are net non-resident purchases of bonds from balance of payments - consistent portfolio flows data. Pull factors are China-specific variables. Push factors are advanced economy (US) variables. All variables, except the lagged independent variable Flows_t-1 ^[1], are measured in time t, reflecting the assumption that investment decisions are dynamic and respond to current events in financial markets. We have chosen robust least squares (RLS) regression (M-estimation) as the estimation method, given that it is less sensitive to outliers in the data set.

For the dependent variable Flows_t, we use monthly data on net non-resident purchases of Chinese bonds (i.e., portfolio debt flows) in billions of USD (Appendix A). As noted earlier, the data are provided by the IIF and collected from national sources such as central banks and national statistics agencies. The series is a proxy for portfolio flows as measured in the balance of payments, yet it has a higher frequency and a shorter publication lag than official quarterly balance of payments data, allowing for a timelier analysis of capital movements. The data set includes 94 observations, from January 2015 to October 2022. Table B2 in Appendix B presents descriptive statistics of the data series.

For the independent variables in the estimation exercise, we consider various alternatives for push and pull factors. Table B3 in Appendix B summarizes the explanatory variable options, and their descriptive statistics are provided in Table B2. Data are obtained via Bloomberg, except in the case of the China policy uncertainty index (see discussion below). The functional forms for all variables are based on standard stationarity tests (Augmented Dickey-Fuller tests).

In line with the literature on common push variables (Koepke, 2018; Koepke, 2018b), we choose a US-based variable to represent advanced economy interest rates. The interest rate variable options considered are monthly percentage point changes in the US Federal Reserve’s Effective Federal Funds Rate and in the US 10-year Treasury yield, as well as three variables that capture monthly changes (in %-points) in expectations for US policy interest rates, calculated from US Federal Funds futures contracts for the subsequent one, two, and three years ahead. For a push variable capturing global risk aversion, we use a monthly percentage point change in the BBB-rated US corporate bond spread over US Treasuries. As an alternative risk aversion variable, we use a monthly percentage change in the VIX index, which is calculated from options contracts to measure expected volatility in the S&P 500 equity index over the following 30 days. Both risk aversion metrics are commonly used in literature studying push variables for portfolio flows into emerging markets (Koepke, 2018b). Figures B1–B7 in the Appendices depict the global push factor options studied.

For the pull variables capturing China’s economic performance, we use a monthly percentage point change in Bloomberg’s consensus forecast for China’s real GDP growth for the following 12 months. As consensus surveys focus on the current and subsequent year’s economic growth, we use a weighting to capture a growth estimate for the next 12 months.^[1] An alternative pull factor measuring economic performance is a monthly average of the Citigroup’s Economic Surprise Index for China. It is a quantitative measure of surprises triggered by economic data releases; when published data exceed analyst expectations surveyed by Bloomberg, the index increases and vice versa. Given the US-China trade conflict of recent years, we also consider a variable that measures economic policy uncertainty in China, as changes in the policy environment may contribute to investors’ willingness to invest in Chinese securities. The Economic Policy Uncertainty Index for China,^[2] developed by Baker et al. (2016), is a news-based indicator that tracks policy-uncertainty-related key words in the South China Morning Post, Hong Kong’s leading English-language newspaper. To our knowledge, this data series has not been previously utilized in the literature on push-pull factors. The variable we use is a monthly percentage change of the index; an increase implies higher policy uncertainty.

In addition, we consider three different Chinese asset return indicators. The first is a monthly percentage change in the MSCI China Index,^[1] which measures Chinese equity returns. The second is a debt return indicator based on monthly percentage changes in an index by JPMorgan Chase & Co that tracks Chinese government and policy bank bonds.^[1] The third is a monthly percentage point change in the yield of the Chinese government’s benchmark 10-year bond. The aforementioned Chinese pull factor options are portrayed in Figures B8–B13 in the Appendices.

3.2 Static Coefficient of Push and Pull Factors in China’s Debt Flows

We begin the process of configuring China’s empirical bond flow model with a specification that mimics the theoretical version presented above. Therefore, the explanatory variables included in the estimation of Model 1 are:

1. Pull variable: China’s real GDP growth forecast, GDP forecast_t (a monthly percentage point change in month t in Bloomberg’s analyst consensus forecast for China’s real GDP growth for the next 12-month period);

2. Push variable: the US Federal Reserve’s benchmark interest rate, EFFR_t (a percentage point change in the Effective Federal Funds Rate in month t from the prior month, calculated based on a monthly average of daily data);

3. Push variable: International investor risk aversion, BBB spread_t (a percentage point change in the BBB-rated US corporate bond yield spread over US Treasury securities in month t from the prior month, calculated based on a monthly average of daily data).

Considering the various options for the economic growth, US interest rates, and the risk aversion variables, we chose the ones that are commonly used in recent related literature. The RLS regression results for Model 1 are reported in Table 1 below. In line with expectations and the theoretical model, economic growth prospects in China, changes in US interest rates, and international investor risk aversion have an impact on bond flows into China. Indeed, China’s real GDP growth variable has a positive sign and the variables on US interest rates and risk aversion have negative signs.

To find an alternative specification for the model, we use forward stepwise robust regression. Now, we go through all possible variables that were discussed in Section 3.1 and presented in Table B3 in the Appendix B one by one to determine which work best in explaining China’s debt flow movements, with the aim of being able to expand the basic model. In the stepwise procedure we perform a series of RLS estimations, starting with a version with a constant and the lagged independent variable. At each step, all candidate variables are evaluated based on their z-statistic. The most significant variable is added to the model until none proves significant at at least the 10 % level. The resultant Model 2 includes the following explanatory variables (in addition to a constant and the lagged explanatory variable):

1. Pull variable: China’s real GDP growth forecast for the proceeding 12 months in month t, GDP forecast_t;

2. Pull variable: Chinese equity return, MSCI China_t (a percentage monthly change in month t in the Chinese equity index, MSCI China, calculated based on a monthly average on daily data).

The estimation result for Model 2 is reported in Table 1 above. Interestingly, the stepwise estimation results suggest that China’s bond flows are only driven by China-specific pull factors, i.e., economic growth prospects and equity returns, as none of the push factors are significant enough.

The importance of China-specific pull drivers in the form of Model 2 are likely to reflect the fact that China’s capital account is still relatively closed. Global investors’ China exposure may be a result of their longer-term asset allocation decisions that are made based on China-specific assessments and are therefore less responsive to short-term developments in global markets. Moreover, as China restricts outbound portfolio flows more than inbound flows, it is unlikely that global factors are properly represented in the model. Nevertheless, Model 1 provides evidence that US interest rates and global risk aversion have an impact on driving portfolio flows in and out of China. Based on model comparison and selection criteria for RLS (adjusted R², Akaike information criterion, and Schwarz criterion), Model 2 outperforms Model 1.^[1] The robust statistics are presented in Table 1 above. Model 1 emphasizes the importance of push variables and Model 2 highlights pull variables. Further discussion is provided in Section 3.3.

Given that both models show that changes in China’s real GDP growth forecasts trigger significant portfolio flows moves, an important variant to the models is a specification that tests whether the impact of changes in economic growth forecasts is symmetric. For this, we modify both Model 1 and Model 2 by including a dummy variable D1 that is equal to 1 for months that saw economic growth forecasts revised higher, and another dummy variable D2 that is equal to 1 for months when growth forecasts were revised down:

The augmented model’s estimation results are reported in Table 2 below. They suggest that the positive impact of upward revision to China’s economic growth forecast is much more notable than the effect of a downward revision (with coefficients of 22.9 and 14.1, respectively, for Model 1, and 19.1 and 10.3 for Model 2). When economic growth forecasts are revised down, China still attracts net debt inflows, yet on a much smaller scale than after positive revisions. Further study of such dynamics may show that China’s restrictions on capital outflows lessen the impact of negative news. Another possibility is that negative economic news dampens intentions of international investors to increase their portfolios’ China exposure. In the augmented Model 1 and Model 2, the variable capturing upward revisions remains highly significant, while the downward counterpart remains significant at 0.10 level in Model 1 and becomes insignificant in Model 2. Robust statistics (presented in Table 2) imply that the outperformance of the augmented Model 2 is statistically significant.

3.3 Extending the Model with Time-Varying Coefficients of Push and Pull Factors

Theory and existing empirical literature on push and pull drives of capital flows to emerging market economies suggest that both push and pull factors should be included in China’s bond flow model. Nevertheless, Models 1 and 2 result in partially contradicting results regarding the drivers of flows. This likely reflects the fact that China has incrementally relaxed capital controls in recent years, particularly for inbound flows, that triggered changes in the significance of capital flow drivers. Increasing financial openness likely makes Chinese bond flows more susceptible to global push factors. Accordingly, when it is assumed that the drivers remained constant through the entire sample period from 2015 to 2022, the results are sensitive to the chosen variable selection techniques. Financial openness is likely to impact the significance and magnitude of the push and pull factors.

To better understand the dynamics of China’s debt flows and their push and pull factors, we extend the study to time-varying coefficients and use the rolling RLS method. The method allows us to use updated infromation on key push and pull variables as time goes on. To accommodate time-varying features, the model can be specified as follows:

The effect of the explanatory variables on month t is represented by β_it for a pull variable i and by γ_it for a push variable i. We use the same explanatory variables as in Model 1, i.e., the lagged dependent variable Flows_t-1 well as GDP forecast_t (pull factor), EFFR_t (push factor), and BBB spread_t (push factor). We use a window size of 40^[1] and a step size of one month. Unfortunately, there is no theoretical guidance regarding the optimal window length in rolling estimation. Indeed, choosing the window size is a trade-off between biases and variance^[2]; using the earliest data irrelevant to the present data-generation process may result in a higher bias, while reducing the sample size causes an increase in variance. Figures B14–B19 in the Appendices present the time-varying rolling coefficients with their 95 % confidence bands and rolling R². The results show that the determinants of debt flows change considerably over time. All variable coefficients display time variation and the notable changes in them coincide with significant global developments.

The global events of 2018–2020 were exceptional, and likely contributed to the instability in the regression model implied by the behavior of the regression’s rolling coefficients, p-values, and R²-values. Since 2020, the real GDP variable has been a notable and stable driver of flows. Indeed, improved economic growth prospects in China encourage international investors to add China-exposure into their portfolios. Meanwhile, the coefficients of the US interest rate and risk aversion variables have steadily become more negative, implying that global push factor have recently become more important drivers of debt flows. As such, higher safe-haven returns lessen the relative attractiveness of China as an investment destination. The result suggests that increased investor risk aversion triggers more notable flight-to-safety activity. While we note that unique global events in the early years of the sample likely reduced the model’s explanatory power and caused significant variation in the coefficients of the explanatory variables, the recent improvements could also reflect China’s capital account opening measures and bond market developments of recent years. The Northbound Bond Connect has made the drivers of flows somewhat more global in nature since its launch in July 2017, and given the 40-month rolling window, the impact of the Bond Connect program is fully accounted for in the estimation from November 2020 onwards. Similarly, the inclusion of Chinese bonds in Bloomberg’s and JP Morgan Chase & Co.’s bond indices began in April 2019 and February 2020, respectively, and were concluded in November-December 2020, after which the negative coefficients of the global push factors have been larger and statistically more significant.

The RLS regression results with time-varying coefficients of the independent variables show a significant improvement in the model stability and explanatory power since early 2020. The timing implies that the launch of the Northbound Bond Connect in July 2017 is a relevant factor explaining the results. China’s real GDP forecast variable has since become highly significant, yet its importance as a driver of flows has decreased. The risk aversion and the Fed’s interest rate variables become statistically more significant following to the trade war and the first wave of the COVID-19 pandemic, with their coefficients shifting deeper into negative territory.