Capital Flow Data - A Guide for Empirical Analysis and Real-time Tracking (2024)

A. Key Capital Flow Data Sources

Table 1 presents detailed information on data characteristics as well as advantages and caveats relating to most commonly used data sources among scholars and practitioners.⁴

B. Balance of Payments Framework: Common Misconceptions and Measurement Limitations

Navigating capital flow data is more difficult than most other economic data, resulting in several common misconceptions and frequent miscommunication among economists (see, for example, the discussion in Avdjiev et al. 2020). This is in part due to the two-way nature of international capital flows: Inward investment reflects changes in foreign holdings of domestic assets, while outward investment reflects changes in domestic holdings of foreign assets. Both inward and outward investment flows can be positive or negative. And the two can be netted against each other to obtain a “net” measure of capital flows.

To provide clarity on common misconceptions, it is useful to illustrate how capital flows are captured in the financial account of the balance of payments, shown in Figure 6. In this illustration, changes in financial assets are separated from changes in liabilities. For each of the two, capital flows can be broken down further into the main components (FDI, portfolio equity, portfolio debt, and other investment).¹⁴

There appear to be three main sources of misunderstandings, corresponding to the different levels of Figure 6:

• First, there are several relevant levels of netting, making the commonly used terms of “net” and “gross” capital flows ambiguous. For example, an important measure of capital flows is the net change in liabilities in the financial account, which captures the net acquisition of a country’s assets by non-residents. This is a “net” measure of capital flows in the sense that different liability components (inward FDI, inward portfolio flows, etc.) are netted against each other. But it is a “gross” measure of capital flows in the sense that changes in liabilities are not netted against changes in assets. Indeed, much of the literature refer to this concept as “gross capital flows” (e.g., Chuhan, Claessens and Mamingi 1998; Broner et al. 2013; Forbes and Warnock 2012; see also the discussion in Avdjiev et al. 2020).

• Second, long-standing sign conventions changed with the introduction of the sixth edition of the IMF’s Balance of Payments and International Investment Position Manual (BPM6) in 2009. The most notable changes included that an increase in foreign assets (including reserves) is shown with a positive sign (and the same holds for a decrease in liabilities; see Table 2 below).

• Third, there are several different types (or “components”) of capital flows. The main components are foreign direct investment (FDI), portfolio equity and debt flows, and other investment (which mainly captures banks’ deposit and lending transactions). These different types of capital flows differ greatly from each other in their empirical determinants and dynamic behavior (Koepke 2019). Yet, the umbrella term “capital flows” is commonly used to refer to a particular type of capital flows, notably the more volatile portfolio and bank-related flows.¹⁵ Moreover, the term “capital flows” is sometimes used loosely to refer to net purchases of investment fund shares (more appropriately referred to as “fund flows”).

A. Meta-Study of Data Sources Used in the Empirical Literature

In this section, we conduct a meta-analysis of the use of different capital flow datasets in the empirical literature. We build on the work of Koepke (2019), which surveys 34 empirical studies to analyze the empirical drivers of capital flows to emerging markets. We extend this sample by including an additional 54 studies, focusing on papers with a large number of citations and/or published in leading economic journals. We evaluate a total of 88 studies published over the course of the last 27 years and listed in Annex III.¹⁷

Sixty-three studies in our sample (or about 70 percent) were published after 2007, likely reflecting increased scholarly interest in capital flow issues as a result of the global financial crisis (GFC) and the ensuing policy challenges. A breakdown by data frequency shows that data used up until the GFC was predominantly at the annual frequency (around 50 percent of studies published between 1993 to 2007), followed by a shift towards much greater use of higher data frequencies in studies published since 2008. Studies using annual data only account for around 20 percent of papers in our sample in the 2008 to 2019 period, with quarterly data around 50 percent and monthly or higher data frequencies around 25 percent.¹⁸

The growing focus on high-frequency data is mirrored in greater use of portfolio flow data, given that for many countries FDI or bank flow data are not available at frequencies greater than quarterly data (Figure 8. and Figure 9. above). One benefit of high-frequency data for empirical research is to facilitate event studies (e.g., Fratzscher 2012; Fratzscher et al. 2012) Among the papers with monthly, weekly, or daily data, the majority use fund flow data, notably EPFR. There are only a few studies that make use of BoP-consistent portfolio flows at a high data frequency (e.g., Converse 2018; Koepke 2018; Cerdeiro and Komaromi 2019; Figure 10. below¹⁹).

The dominance of fund flow data in recent academic research raises an important concern about the validity of academic findings for “true” capital flow data. For example, recent research has emphasized the importance of external (“push”) factors for portfolio flows (e.g., Feroli et al. 2014; Friedrich and Guerin 2019). However, fund flow data by construction are more affected by push factors, given the role of investment benchmarks in the portfolio allocation process (see, for example, Raddatz et al. 2017; Arslanalp et al. 2020). Therefore, using fund flows to examine the role of push versus pull factors is likely to bias empirical findings towards greater importance of push factors.

The use of different data frequencies itself may also affect the estimated statistical significance and economic importance of push versus pull factors. For example, Ananchotikul and Zhang (2014) find that at short time horizons, push factors play a dominant role in driving fund flows but at longer time horizons pull factors become more important as weekly fluctuations in push factors cancel out.²⁰

We investigate this issue further by classifying papers in our sample according to whether they study push factors, pull factors, or both (Figure 11). We find that studies using high-frequency data generally tend to focus more on external factors while studies using lower frequency data tend to focus more on domestic factors. This pattern is consistent with the idea that the relative importance of push vs. pull factors may be partly determined by the data frequency.²¹

B. Constructing a Monthly Portfolio Flow Dataset Consistent with BoP Principles

To help address the potential biases stemming from the use of BoP-inconsistent portfolio flow proxies, we construct a dataset on monthly emerging market portfolio flows that is broadly consistent with balance of payments accounting principles and specifically geared towards academic use.²² The dataset is similar to the monthly data on country-level portfolio flows compiled by the IIF, but has the advantage that it is available online for free (whereas IIF data are not accessible to most academics because they are restricted to members, which are mostly financial institutions). The dataset is posted online along with this paper and will be updated periodically (Link).

In the rest of this paper, we refer to this dataset as the KP dataset (short for Koepke & Paetzold). The KP dataset is constructed using data from national sources in a set of 19 EMs²³, with data for aggregate portfolio flows for 19 countries and debt and equity flows for 18 countries. Data availability begins in 2010 for most countries. Figure 12 shows total monthly equity and debt flows, respectively.

Of the 19 countries, 13 report monthly data that correspond to the quarterly balance of payments published by the central bank or statistical authority. For six countries, we collect monthly proxies for one or more portfolio flow type.²⁴ Table 3 provides additional details on the data used in the KP dataset.

Figure 13 show total EM equity and debt flows, respectively, for the KP data and three relevant comparators (IIF monthly tracker, IIF daily, EPFR monthly). The sum of total flows to EMs in the KP dataset from 2010:Q1 to 2019:Q2 covers around 45 percent of total flows in the BOPS, with somewhat higher coverage of equity flows (55 percent) and lower coverage for debt flows (42 percent). Equity data from the KP dataset and the IIF’s monthly portfolio flow tracker are highly correlated with quarterly BoP-based equity flows (see Table 4). For the debt component, the KP data has a high correlation with BoP flows at the country level, but somewhat lower at an EM aggregate level.

Figures 14 and 15 illustrate how the KP data compare to balance of payments data. Average quarterly flows in 18 country cases, with the exception of South Africa, are equal or close to averages from the balance of payments. Co-movement for the majority of countries in our dataset with available data shows a standard deviation roughly in line with the standard deviation of balance of payments data, with the exception of Mexico, Sri Lanka, South

Africa, and Thailand.²⁵ In comparison to investment fund data such as EPFR, the dataset provided in this paper more closely aligns with standard definitions in the BoP. Using this dataset may help researchers obtain empirical findings that are valid for “true” capital flow data, consistent with established BoP accounting principles.

Nowcasting Horse Race

We use several alternative portfolio flow proxies in a nowcasting setting with quarterly portfolio flows from the BoP as the dependent variable, where the accuracy of nowcasts is measured as the root mean squared forecast error (RMSFE). The portfolio flow proxies include:

• Weekly EPFR data, published with a lag of about 7 days.

• Monthly EPFR data, published with a lag of about 15 days.

• Daily IIF data, published with a lag of 1–3 days.

• Monthly IIF data, published with a lag of about 1–5 days.

We also include the monthly data from the KP dataset, described in the prior section. The underlying data are typically available after about 2–3 months.

The sample period for the nowcasting exercise runs from 2010:Q1 until 2019:Q2. In our baseline model, we use aggregate flows for those emerging market countries where all the portfolio flow proxies are available (separately for equity and debt flows).²⁸ Aggregate flows to these countries are scaled by their combined GDP. We estimate a pseudo-out-of-sample forecast error, using only data that would have been available as of the relevant quarter. Therefore, the number of observations used in the regression is quite small at the beginning of the sample period and increases over time.

In the first step, observations for each portfolio flow proxy are summed up incrementally for each quarter. For example, daily capital flows are summed for days 1–2, days 1–3, days 1–4, and so forth, up until the last observation of the quarter is included. Similarly, weekly/monthly data are summed over an increasing number of weeks/months in each quarter.

In the second step, for each of these sums the following model is estimated:

BoP_quart = c + β_ij · capflow_sum_ij

Where BoP_quart is quarterly portfolio flow data from the balance of payments, c is a constant term, and capflow_sum_i is the sum of the high-frequency proxies for portfolio flows from the first observation to the i-th oberservation of the quarter, and j represents the portfolio flow proxy used.

This model is used to predict quarterly BoP portfolio flows using available data from 2010:Q1 onwards. For example, to predict the quarterly BoP flows in 2012:Q1, the model is estimated using the observations for 2010:Q1–2011:Q4, while for the prediction for 2012:Q2, the model is estimated using the observations for 2010:Q1–2012:Q1. For each quarter, i nowcasts are produced by multiplying the estimated coefficient from the regression and capflow_sum_i for that quarter (and adding the constant term).

We evaluate the predictive content of each portfolio flow proxy by computing the root-mean squared forecast error for each i, for the period from 2012:Q1 until 2019:Q2:

$RMSF{E}_{ij}=\sqrt{\frac{1}{T}\mathrm{\Sigma }(y_{tij}-{\hat{y}}_{tij}{)}^2}$

Where RMSFE is the root-mean squared forecast error for sum i and proxy j, T is the number of quarters for which a forecast (or “nowcast”) is produced, t is the time in quarters, y is actual portfolio flows for that quarter, and ŷ is the forecast for that quarter.

The results of this exercise are shown in Figure 16 for equity flows and in Figure 17 for debt flows. As expected, forecast errors for all proxies generally decline over the course of the quarter as more data become available. Daily IIF data and weekly EPFR data are available as early as 2 and 7 days into the quarter, much earlier than either of the monthly proxies. RMSFEs based on the first few observations of daily and weekly proxies are relatively high, but the forecast performance of the daily IIF data improves notably once 20–30 days’ worth of data are included.

The monthly proxies first become available about 33 and 45 days into the quarter for IIF and EPFR data, respectively. The IIF’s monthly tracker consistently performs better than either of the EPFR datasets (weekly or monthly), but only outperforms the IIF daily flows data after the second data release in the quarter (on average). Similarly, EPFR’s monthly data mostly underperforms the weekly data, with the exception of debt flows towards the end of the quarter. We also compare the RMSFEs of the various proxies to that of a benchmark autoregressive regression with one lag and a constant term. All proxies outperform this benchmark by a wide margin, with all RMSFEs at least 80 percent lower.

The improvements in forecast performance of quarterly capital flows using IIF compared to EPFR data are economically important. Expressed as a share of absolute average quarterly flows, RMSFEs of debt and equity flows based on daily and monthly IIF data outperform weekly and monthly EPFR data, respectively, by around 20 percent early in the quarter and 50 percent at the end of the quarter. Compared to the benchmark regression, RMSFEs of IIF flows as a share of quarterly flows are consistently lower while EPFR data only achieve marginally lower RMSFEs at the end of the quarter.

The results also show that the KP dataset has a similar performance to the IIF’s monthly tracker, although the monthly data from countries contained in the KP dataset is subject to a greater release lag within the current quarter. One caveat regarding the IIF’s monthly EM tracker and the KP dataset is that observations for the latest 1–2 months of data are revised with each monthly data release for the EM aggregate estimate and for more than half of the countries included in the country level flows. This may bias our results towards a smaller forecast error for these datasets. The other data sources generally do not get revised.

The main findings from the nowcasting horse race are that (1) all portfolio flow proxies considered have significant predictive content for BoP portfolio flows, (2) IIF proxies generally outperform EPFR proxies, and (3) higher frequency proxies generally outperform lower-frequency proxies, especially in the first half of the “current” quarter.

We test the robustness of these findings through a range of alternative specifications. For example, we conduct the same exercise using a longer sample period that begins in 2005:Q1, which did not substantially change average forecast errors. In an extension inspired by the coincident indicators proposed by Miao and Pant (2012), we include averages from the first to the i-th quarterly observation of global financial variables used in the literature on capital flow drivers (VIX, US Treasury yields) as additional predictors. The inclusion of these variables does not consistently reduce the RMSFEs and the relative magnitude of forecast errors does not change substantively.

We also run the nowcasting exercise using aggregate EM flows from each source (for equity flows, debt flows, and the sum of equity and debt flows), as well as for individual EM countries where data are available from all three data sources. The results are presented in Annex IV. For the EM aggregate exercise, the monthly proxy data often perform better than in our baseline results. Notably, the IIF’s monthly tracker generally outperforms other indicators for both total flows and equity flows, while EPFR tends to outperform for debt flows.

1

We are grateful for helpful suggestions by Gustavo Adler, Swarnali Ahmed, Serkan Arslanalp, Mahir Binici, Robin Brooks, Eugenio Cerutti, Nathan Converse, Dimitris Drakopoulos, Thomas Elkjaer, Eric Fischer, Jonathan Fortun, Joji Ishikawa, Sebnem Kalemli-Özcan, Yevgeniya Korniyenko, Swapan-Kumar Pradhan, Gian Maria Milesi-Ferretti, Gurnain Pasricha, Daniel Rodriguez, Russell Rollow, Todd Schneider, and Ling Zhu. We would also like to acknowledge the valuable comments from participants in several IMF seminars on monitoring capital flows, including during presentations given to the IMF’s Capital Flows Group, the Asia & Pacific Department, the Western Hemisphere Department, and two ICD lectures on “Tracking, Interpreting and Forecasting Capital Flows Developments.”

1

The numbers referenced for IIF and IMF data relate to the concept of “gross” portfolio inflows, defined as non-resident purchases of EM stocks and bonds. ^{Section 2} provides a detailed discussion of differences in scope and data coverage of BoP, EPFR and IIF data.

2

An important benefit of using capital flow data on a BoP basis is that they are consistent with other macroeconomic accounts, such as current account data, government finance statistics, and the international investment position. Capital flow data are often analyzed through the prism of external financing (of a current account deficit) and financing of a government’s budget deficit. Indeed, a key motivation of the voluminous literature on the role of external (“push”) and domestic (“pull”) drivers of capital flows has been to advance our understanding of the degree to which foreign capital inflows are likely to be a temporary rather than permanent source of external finance (e.g., Calvo et al. 1993, Chuhan et al. 1993; Fernandez-Arias 1996; surveyed in Koepke 2019). This motivation argues for the use of BoP-consistent data that can be analyzed and interpreted in conjunction with other standard macroeconomic data.

3

For the academic dataset, we build on the work of Koepke (2019), which surveys 34 empirical studies to analyze the empirical drivers of capital flows to emerging markets. We extend this sample of studies by including an additional 54 studies in this same strand of literature, focusing on papers with a large number of citations and/or published in leading economic journals. Overall, we evaluate a total of 88 studies published over the course of the last 27 years and listed in Annex III. For the policy dataset, we review 220 reports published since 2010, including financial stability reports by G-20 emerging market central banks and reports of policy-oriented international financial institutions, notably the IMF’s Global Financial Stability Reports, the BIS’s Annual Reports, and the World Bank’s Global Development Prospects report.

4

The use of capital flow data sources in the academic literature is discussed in detail in section 3. Data sources commonly used by practitioners for tracking capital flows in real time are discussed in section 4.

5

The BPM6 is available at https://www.imf.org/external/pubs/ft/bop/2007/pdf/bpm6.pdf. The BPM6 compilation guide is available at https://www.imf.org/external/pubs/ft/bop/2014/pdf/guide.pdf.

6

The BPM6 notes that “residence of each institutional unit is the economic territory with which it has the strongest connection, expressed as its center of predominant economic interest” (¶4.113).

7

It is worth noting that the term “flow” has a different meaning in the capital flow literature than in the terminology of balance of payments statistics. In the capital flow literature, a “flow” relates to transactions. In macroeconomic statistics, a “flow” refers to all changes in the stock of a variable from one period to the next, including transactions, valuation changes, and other flows (see BPM6 ¶2.2).

8

The BPM6 specifies that “a single transaction between two counterparties thus gives rise to four entries. […] As a liability of one unit is mirrored in a financial asset of another unit, for instance, they should be identically valued, allocated in time, and classified to avoid inconsistencies in aggregating balance sheets of units into regional or global totals” (¶3.29).

9

The BPM6 notes that “market prices are the basis for valuation in the international accounts” (¶3.67). However, some financial instruments (e.g., loans) are recorded at nominal values.

11

Pradhan and Silva (2019) show that the data contained in the LBS align relatively closely with the IMF’s International Investment Position Statistics but there are several sources of discrepancies, including differences in reporting populations, mix of data sources, treatment of instruments, and country-specific factors.

12

The LBS provide bilateral data for both claims and liabilities whereas the CBS provide data mainly for claims and very limited data on liabilities (overall total and local currency liabilities in host countries).

13

More specifically, the data serve as a proxy for non-residents’ net acquisition of EM assets, i.e. inward portfolio flows to EMs. For some countries, the sum of daily or monthly observations is identical to the official quarterly BoP data on inward portfolio flows for the same period.

14

The financial account also records financial derivates, but these are excluded in this discussion of capital flow data.

15

For example, the paper entitled “Capital Flows are Fickle: Anytime, Anywhere” (Bluedorn et al. 2013) mainly focuses on volatile components of capital flows. Similarly, a recent paper on “Capital Flows at Risk” (Gelos et al. 2019) focuses on portfolio equity and debt flows. The use of the term “capital flows” is thus similar to the use of the word “America” for the United States of America (a stylistic device referred to as totum pro parte, the whole for a part).

16

The study includes the following countries with high-quality fund holding data: Australia, Canada, Denmark, European Monetary Union, Norway, Sweden, Switzerland, the United States, and the United Kingdom.

17

Our sample begins in 1993 with the seminal study by Calvo, Leiderman and Reinhart. In our selection of papers, we give preference to studies with 20+ citations and those published in the top two quartiles of economic journals per the ranking by SCImago Institutions Rankings. For recently authored studies, we err on the side of including additional studies because citations will tend to increase with time.

18

Information on data frequency or data source was unavailable or ambiguous in some papers such that the percentages do not sum to 100 percent.

19

By BoP consistency we refer to data that broadly follows balance of payments accounting principles.

20

One implication of this effect could be that other capital flow components may also be more subject to common external forces at short time horizons. In the absence of high frequency data on bank flows and FDI, it is not clear whether this hypothesis can be tested empirically.

21

In contrast to other capital flow components, bank flows have rarely been studied at the annual frequency, given that the BIS provides quarterly data, which provides the benefit of a greater number of observations for quantitative analysis.

22

In particular, the data are based on transactions between residents and non-residents of emerging market countries. The data are not subject to valuation effects and reflect transactions at market prices.

23

The countries included are the Brazil, Bulgaria, Chile, China, Czech Republic, Hungary, India, Lebanon, Mexico, Pakistan, Philippines, Poland, Romania, South Africa, Korea, Sri Lanka, Thailand, Turkey, Ukraine. Of these countries, the Czech Republic and Korea are not part of the IMF’s classification of emerging markets, but are included in private sector classifications of EMs such as leading investment benchmark indices. Data for Bulgaria, Hungary, India, Romania, and South Africa were converted to USD using monthly average exchange rates.

24

Among the proxy data, we use government debt flows for Mexico, non-government debt flows for Thailand, transactions on national stock exchanges by non-residents for South Africa and Sri Lanka, portfolio inflows by non-residents for India and total non-banking sectors’ portfolio flows for China. Only total monthly flows are available for China.

25

Potential reasons for these discrepancies could be data revisions as well as the use of proxy data and incomplete data for capital flows (i.e., stock exchange data in the case of equity flows for Sri Lanka and South Africa, the exclusion of government debt flows for Thailand, and the exclusion of private sector debt flows for Mexico and Sri Lanka).

26

In this section, the terms “nowcast,” “forecast,” and “prediction” are used interchangeably.

27

As noted in Miao and Pant (2012), variants of this approach have been used by several studies to obtain timely proxies of capital flow movements, including Calvo et al. (2004 and 2008). One limitation of this approach is that trade in services, primary income, and secondary income play major roles in some EMs, causing errors in the estimation. The IIF provides monthly estimates of net capital flows for 23 EMs, using the same basic approach.

28

For equity flows, the sample contains Brazil, India, Pakistan, Philippines, South Africa, Korea, Sri Lanka, Thailand, and Turkey. For debt flows, the sample contains Hungary, India, Poland, South Africa, Thailand and Turkey.

1

Country A’s outward direct investment position is also derived using all participating countries’ inward direct investment positions from country A as a proxy for country A’s outward direct investment position.