Re-Emerging Markets

Introduction

One key issue is whether the stylized empirical facts about emerging markets may be due to selection bias. In particular, Harvey (1995) notes that the high means may be partially due to the survival of the emerging market. Analytical results and simulations in Brown, Goetzmann and Ross (1995) [BGR] suggest that in the presence of a performance threshold, high ex post means are a direct consequence of survival. They further find that bias in the mean is functionally related to the variance of the series. These results have potential implications for the analysis of emerging markets, since we only observe markets that have recently emerged above a threshold. In contrast, submerged markets have fallen, or remained, below the threshold.

In this paper, we investigate the properties of a group of markets where returns are measured after recent "emergence." Previous work on survival provides analytical solutions in situations where markets cease to exist whenever some index falls below a minimum value. This is different from the situation where markets are created once in a while, but only retain the attention of global investors when their current capitalization is large enough. We simulate a simple, general model of global markets, in which some are likely to emerge early and some are likely to emerge late. Our simulations indicate that conditioning upon market history, where history is defined as the length of time since the market has "emerged" will result in a number of empirical regularities. Among other things, the brevity of a market "history" is closely related to the bias in annual returns imparted by survivorship, as well as to the low level of covariance with the rest of the markets. These findings are potentially useful for international investors, because merely knowing that a market has recently emerged contains information about the future distribution, as well as about its future prospects for survival.

In order to explore the effects of survival upon a universe of world markets we assume that a market is only observed by the econometrician after its last "emergence" above a certain threshold of capitalization. This is the typical emerging market. A market cannot be observed unless it has emerged -- that is, it has gotten large enough, and/or organized enough for the International Finance Corporation (IFC) (or earlier, the League of Nations) to begin collecting data and creating an index. Alternately, the returns themselves may have attracted attention of international investors, who in turn begin to collect information about the markets performance. In the situation in which market emergence is conditioned upon capitalization exceeding a certain trigger level, the expected price path can be solved analytically.² We know, for instance, that when markets are near the lower bound, the bias in returns is greatest, because the probability of hitting the lower bound is high, yet we only observe markets that remain above the threshold. Unfortunately, for more complex forms of survival and ex post conditioning, we can only learn about the expected price path and other variables of interest through simulation.

This paper is organized as follows. Section I illustrates the type of bias that can result from only considering markets that have emerged. Section II presents the results from the simulation model and Section III discusses the results. In Section IV, we provide performance results for markets which have emerged in the last twenty years. Section V contains some concluding comments.

I. Modeling Emergence

Suppose a number of markets started trading 100 year ago, each with about the same capitalization (without loss of generality), but with differing expected returns. This is an over-simplified model, since each market will not only have its own rate of return but its own variance, its own particular production "mix" and its own legal and political environment. In fact, all of these are likely to be nonstationary -- an important issue to be discussed later in the paper. In our simple setting, however, on average, those markets with lower expected returns will have lower capitalization in the future than those with higher expected returns. Under most conditions, time will separate the markets according to their respective drift processes. Figure 1 illustrates the model.

We define a market as an emerged market if it has crossed the barrier at 1 from below at least once since t = 0, and if the last crossing, as of t=100 was from below. In another situation, where markets all start at the same time, but are discarded when they become too small, the properties of the expected price path are like those studied in BGR. In the case where the drift is known and the barrier is constant, the authors show how the conditional mean is biased upward.

The survival bias in BGR is not the whole story, however. The fact that the market crossed the barrier recently provides additional information for estimation of the mean. A market that crossed recently (i.e. later than other markets) is likely to have a low expected return. Thus, historical information about when the market began, what its past capitalization was when it started, and what its capitalization was when it last emerged may all be useful inputs to estimating the actual expected return. The "recentness" of emergence is inversely related to its drift. The conditioning process in the BGR analysis required that processes were "memoryless" since the time of emergence. In the current analysis the long-term memory of when and where the process began is crucial.

I.2 A Stylized Model

In order to explore the empirical implications we now formalize the above argument. As Harvey (1995) shows empirically, the worlds markets differ in their systematic as well as unsystematic risk components. For simplicity, consider a single-factor model excess log return generating process for the equity market for country k,

In this paper, we consider the effect of a capitalization threshold on the observability of returns. The issue is, What would happen to statistical inference about the world stock markets if the returns to markets whose capitalization fell or stayed below a given threshold were unobservable -- i.e. we failed to observe and measure returns in really small markets?

I.3 Simulation Experiment

We simulate emergence among a group of markets as follows.

1. Annual returns for each simulated market are generated by the model in equation (1).

2. We simulate 100-year histories of the global market using i.i.d. normal returns with an annual mean of 10% and an annual standard deviation of 20%.

3. We simulate the local factor, e _k,t , with i.i.d. draws from a normal distribution with a mean of 0 and a random standard deviation. For each market, the annual standard deviation of the residual term e _k,t is drawn from a uniform distribution between 10% and 30%.

4. ß _k, the loading on the global market index for each market, is drawn randomly from a uniform distribution between 0 and 2.

5. Markets begin randomly with starting dates drawn from a random uniform distribution over the interval 1 to 99.

6. All markets start at one standard deviation of r_t below the capitalization threshold.

7. We construct capital appreciation returns for each market, assuming no dividend payments.

8. We cumulate each index from its inception until the last period, to create stock market indices for 100 markets.

9. We "censor" the markets by dropping those which are below the threshold at the terminal date.

Both beta and the residual risk differ to make the model more realistic, as we would expect markets to differ in terms of expected returns and risk. While the extension to a multiple-factor global model is straightforward, the single factor model is used here for simplicity. This procedure is repeated 2,500 times.

For each simulation, we save a number of variables, including the year the market began, the year it last emerged, conditional and unconditional summary statistics; conditional statistics are only for after the last emergence. Of particular interest to emerging market investors is the difference between the mean annual return of the series since the period of last emergence less the mean annual return of the series using all observations. We define this as the "bias" in the annual mean return. We also wish to examine the relationship between beta and emergence, R² and emergence, and finally, the additional information provided by the start date of the series in relation to the emergence date of the series. This last issue is important because it may be possible to develop heuristics for correcting the bias in the mean returns for recently emerged markets. One potentially useful piece of information for this purpose is the length of time that the market has existed. For example, given any two markets that emerged ten years ago, it seems likely that the market with a longer history will have a lower mean, and therefore a greater bias in the mean than the market with a shorter history. We examine this issue via simulation results.

II. Simulation Results

The results of the simulations reveal three important implications of recent emergence. First, the more recent the emergence of the market, the higher the bias in the mean return. Figure 2 shows the relationship between the bias in the mean and the time of last emergence. The figure shows fractiles of the distribution of the ex post mean since last emergence less the actual mean calculated over the entire history of the series. Notice that for early emergence, that is, for markets that emerged in the early part of our hypothetical century, there is virtually no bias in the distributions -- the difference between the sample average annual return and the ex ante returns is typically zero. For very recent emergence, that is for markets which last emerged only five years ago, the bias is as high as 10% per year, with values over 20% per year not unusual. Even for markets that emerged a decade ago, we still see a substantial positive bias in returns, with average bias of about 5% annually.

Second, we find that recent emergence is negatively correlated to the observed loading on the global market factor. Figure 3 shows the average beta of each market, sorted on the year of its last emergence. Markets that emerged near the beginning of the hypothetical century have an average beta of 1.4, while recently emerged markets have much lower average beta, around 1.0. Thus, the period of last emergence is informative about the unconditional expected return of the market -- the more recent the last emergence, the lower the beta and consequently the lower the unconditional expected return. ⁴ Our simulations reveal a third empirical regularity of interest to researchers investigating emerging markets. Figure 4 shows the R-square from the market regression, sorted on period of last emergence. For markets that emerged early, the amount of variance explained by the market is high, about 60%. For markets that only recently emerged, the R-square is under 40%. Thus, by conditioning upon recent emergence, we should expect to find markets with higher ratios of idiosyncratic risk. This corresponds exactly to the observation that emerging markets seem to have low correlations with worldwide factors.

II.2 Bias and Idiosyncratic Risk

The three empirical regularities revealed by the simulations are partly explained by the analytics in BGR. By conditioning the observation of the series upon a lower bound, we are implicitly conditioning upon series survival. For instance, consider an "submerged" market that began a decade ago. We typically do not observe those series that emerged at the same time but failed to finish above the lower bound⁵. This survival conditioning will bias the ex post observed return. BGR find that the bias is greater when the price level is near the lower bound. In the context of our simulations, recent emergence implies a proximity to the bound and thus a more acute survival bias.⁶

BGR also show that the bias increases in the residual variance of the series. We can examine this equivalence directly by looking at the relationship between idiosyncratic risk and return bias, conditional upon the date of last emergence. Figure 5 displays the result of a set of regressions of the bias in the mean upon the residual standard deviation of the series. The regression is performed for each year of last emergence in the sample. We hold this year constant since this is a primary effect. Our regressions confirm that the bias is positively related to the residual risk of the series. Further, the effect is greatest for recently emerged markets. Also reported in the figure are the t-statistics associated with each regression. They are typically over two for the last twenty years of the analysis. This provides an interesting possible application. For two markets emerging at the same time, we expect the one with the highest residual variance to have the largest bias. Empirically, this relationship between bias and residual risk will appear as though the local market factor is "priced," i.e. that the expected return in positively related to the portion of the variance not correlated to global factors.

II.3 Bias and Start Date

The BGR survival results, however, explain only part of the empirical regularities we find in simulation. By conditioning only upon the last emergence, we are allowing markets to emerge and then submerge. Markets that remain in the neighborhood of the boundary over time are also likely to be those that emerged and submerged repeatedly. When we condition upon having emerged early in the hypothetical century, we implicitly "pick out" those markets with positive drift. This in turn decreases the probability of being near the barrier for any length of time. This effect provides another possibly useful heuristic for adjusting mean bias for emerging markets. For any two markets that emerged at the same time, the one that began earlier should have the largest bias. The reason for this is that, if it began a long time previously and still only recently crossed the barrier, it probably has a low beta, whereas a newly formed market that quickly crosses the barrier is likely to have a high beta, and thus a high expected return. Figure 6 shows precisely this effect. For each period we select all the markets that last emerged at that time. We regress the bias in the mean upon the starting date for the series. We find the relationship negative for all periods. That is, the earlier the starting date, the higher is the bias in the mean. For new markets, the bias is lower.

This "sorting" of markets by their expected returns via conditioning upon time since last emergence is potentially useful because investigators have information about when markets existed in the past, and when they last emerged. Therefore the time since last emergence and the time since the market first began may be used to forecast expected returns⁷.

II.4 Cumulative Average Returns: Emergence as an Event

Another approach to the bias issue is to treat the date of emergence as an event and to align the simulations in event time. Figure 7 shows the cumulative and average returns for all markets that last emerged on a given year. Last emergence dates are taken as 10 years, 15 years and 25 years before the end of the simulation period, so the three lines might all be considered representative of "emerging" markets of different vintages. The horizontal axis is aligned in event time, as opposed to calendar time, and market emergence is set to year 0. The figure shows a cumulative index of ten years of returns preceding and following the emergence date.

Notice the strong effect of conditioning upon emergence. Returns are nearly flat before emergence, despite the fact that the average returns in the simulation are positive. Following emergence, the returns follow a positive trajectory, which is slightly is convex. The returns comprising the price path are shown in the second panel of Figure 7. The difference between pre-emergence and post-emergence returns is dramatic. We can reject equality of mean returns before and after emergence with a high level of confidence. ⁸ The low returns preceding emergence are likely to be a consequence of conditioning upon the market being below the capitalization threshold. Simply knowing that a market crossed the threshold from below, and remained above until the end of the sample period helps to differentiate historical returns. Notice, also, that the largest return is in the year immediately following emergence. This is because year 1 is the year in which the market is closest to the boundary, and thus most likely to fail. As the market climbs away from the boundary, the chances of submerging decrease, and the survival-conditioned return decreases as well. If an econometrician only observes the history of a market since emergence, the simulation suggests that the average return will be greater just after emergence.

II.5 Implications

The simulations, however stylized, provide some general guidance for investment practice. First, recent emergence by a market has the potential to be a strong conditioning factor that may affect ex post observed return distributions. Some of the effects are due to the fact that recently emerged markets are by definition near the lower threshold of capitalization. However, some of the effects are also due to actual differences in long-term expected returns, due to differing betas. This sorting also underscores the importance of detecting changing betas for emerging markets. Local or recent changes in expected returns may be sufficient to help a market avoid plunging below the lower threshold. The econometrics of conditional betas (c.f. Harvey, 1995) would appear to be a crucial step in the analysis of future expected market returns.

These simulation results have strong implications for applications of mean-variance optimization to emerging market data. The brevity of emerging market histories induces a well-known uncertainty in the inputs to the mean-variance model, known as "estimation risk".⁹ Our work shows that the problem extends beyond input uncertainty to input bias. Recently emerged markets typically have a positive bias in the mean and wider distribution. In a mean-variance framework, the distribution of the bias is as important as the average, because extreme values exert a large influence upon the composition of the optimal portfolio. Institutional investors seeking data on emerging markets for use in mean-variance optimization should use recently emerged market data with extreme caution. As the number of emerging markets used in the optimization is increased, the likelihood of overweighing one with an extreme positive bias in the mean will increase as well.

III. A Look at History

How well does the history of the global stock markets accord with the central premise of the simulation -- namely that some markets have been around a long time, but only recently have emerged? Table 1 provides a partial list of the founding dates of the world's stock exchanges. It is based upon information in two well-known guides to global stock markets, Park and Agtmael (1993) and O'Conner and Smith (1992). Both of these sources collect information about market histories from currently operating stock exchanges around the world. Because countries that currently have no exchange are not included, this is presumably not a complete list of the markets that existed at one time. However even this partial list is interesting because it tells us just how much we do not know about equity markets. Of the forty markets that were founded before the twentieth century, only two, the U.S. and the U.K. markets, have been extensively analyzed over long investment horizons. This is not from lack of interest, but from lack of data. While econometricians in the U.S. and the U.K. have compiled reliable historical return information stretching back into the ninteenth and eighteenth centuries, it has only been recently that comparable information has become available for other markets such as Germany, France and Switzerland, albeit with notable gaps due to wars. Even so, this table is informative. Table 1 tells us that most of today's stock exchanges have long histories. Many non-European markets began under the aegis of colonial rule, including Hong Kong, India, Pakistan, Sri Lanka, Indonesia, South Africa, Egypt and Singapore, and have continued with or without interruption to the present. Other markets only recently emerged from communist rule -- Hungary, Czechoslovakia, Poland, Romania, Slovenia and Yugoslavia. Perhaps most surprising is the number of South and Central American countries with long market histories. Argentina, Brazil, Colombia, Uruguay, Mexico and Venezuela all have had equity markets for more than sixty years.

The reasons for the disappearance of many of the world's stock markets are well documented. The League of Nations collected data on the capital appreciation of market indices in the period from 1929 through 1942. This collection effort was continued by the United Nations¹⁰. Figure 8, Figure 9 and Figure 10 show 22 monthly real price indices for a number of the world's stock exchanges since 1929. The 1929 starting date is used because this is the earliest monthly inflation data obtainable. Discontinuities in the price indices are handled by assuming that no change occurred. We are currently exploring the possibility of estimating price change over discontinuities, but this involves analysis of the changing composition of the indices. The monthly series (and annual series not pictured) indicate the dramatic fluctuations of many of the world's stock markets in the twentieth century.

While not accurate enough to provided investor return information, the early League of Nations data can be used to assess the broad impact of the Great Depression and World War II on price index continuity. The early price data indicate that the hyperinflation of the 1930s closed the Danish and German markets in the early 30s. While most markets remained open and functional through the Great Depression, World War II caused many of them to shut down in the 1940s. Some Eastern European markets remained open through the war, only to suffer expropriation after 1945. In total, Austria, Belgium, Shanghai, China, Czechoslovakia, France, Hungary, Japan, Korea, Luxembourg, Malaysia, the Netherlands, Norway, Poland, Portugal, Uruguay, Venezuela, Yugoslavia and Slovenia all experienced temporary or permanent shut-downs either in the war years, or in the occupation following the war. Many of the markets which shut during mid-century "re-emerged" after the war or after occupation.

Wars are not the only factors that create discontinuities in the price records. Markets in Egypt, Lebanon, Portugal and Chile, for instance were shut down or barred to foreign investors due to political changes largely uncorrelated to outside global trends. Shifting legal factors have changed the attractiveness of markets such as Greece, Turkey and India to outside investors, and thus caused them to be regarded as "emerging" markets, despite long histories as capital markets. A turn of political fortunes can make a long-forgotten market suddenly of interest to outside investors. A fair question is whether these events should be considered endogenous or exogenous -- clearly some of these formerly emerged countries failed to maintain a social and political system which fostered steady industrial growth. While it is not the object of this paper to ask why these markets did not prosper in the same way that the U.S. and the U.K. markets did over the past sixty years, it is a reasonable to ask whether anything has changed. Have the political and economic forces that caused these markets to submerge been fundamentally altered? Can we expect the next sixty years of capital markets to be different from the last sixty years?

There are a few "emerging" country indices shown in Figure 10: India, Pakistan, Philippines, Venuzuela. While two of them, India and Pakistan have shown drmatic increase since the mid-1980's, neither index is above its peak value acheived around mid-century. Indeed, only one market of the three, India, is currently above its starting index value.

IV. Empirical Analysis of Emerging Markets

One of the obvious problems with evaluating survivorship bias in emerging markets is that the data may not be readily available before markets are considered to have "emerged." Yet a number of empirical regularities should be expected from the simulations. For instance, right after emergence, the bias should be greatest since we know that markets that emerged then subsequently submerged are dropped from the analysis. This hypothesis is analyzed using a variety of approaches.

IV.1 Selection of "Emerged" Markets

The standard data source for emerging stock markets is from the International Finance Corporation (IFC), the private-sector institution that is a sister of the World Bank. In its laudable quest for promoting private equity investment in Less Developed Countries (LDCs), the IFC has collected the most comprehensive and consistent data base for Emerging Markets (EMs). The data base started in December 1980 with 9 markets, which were backfilled to December 1975, and has expanded to 22 markets as of December 1995. Other markets are being watched by the IFC, then periodically added to their composite EM index. The IFC attempts to collect not only share prices, but also dividend information which allows investors to calculate the total return to equity investment in the country over the period covered.

The term "emerging stock market" was coined by the IFC in 1981. IFC defines an emerging stock market as one located in a developing country. Using the World Bank's definition, this includes all countries with a GNP per capita less than $8,625 in 1993. The IFC states that "although IFC has no predetermined criteria for selecting an emerging market for IFC index coverage, in practice most markets added have had at least 30 to 50 listed companies with market capitalizations of US$1 billion or more and annual value traded of US$100 million or more at the start of IFC index coverage." This definition clearly defines a size threshold that markets have to reach before official inclusion in the database.

In practice, two types of survivorship biases are imparted to IFC indices. The first is the selection of the market itself and is the primary focus of this paper. The second affects the construction of some early IFC indices, which were started in 1980 using companies that were in existence at that time. As the series were backfilled to December 1975, an additional bias is due to company selection. In an earlier era, this data collection activity was carried out by another international institution -- the League of Nations statistical service. Beginning in the 1920s the League of Nations recorded share price indices for a number of countries. This data was principally supplied by the central bank of the country, or the stock markets themselves, and included no dividend data. More recently, the International Monetary Fund (IMF) has compiled similar stock index data.

The IFC dataset has become the standard database for research on EMs and provides performance benchmarks for portfolio managers. As a result, the introduction of new markets is watched very closely by portfolio managers given that it will affect the return on their "bogey." We consider that the first date at which the IFC collects data for a market as the data of emergence. Table 2 presents start dates and market capitalizations for the markets covered by the IFC as of January, 1994. In addition, the table indicates the year of the founding of the stock exchange in each country, and the annual price index data available from the League of Nations and International Financial Statistics Yearbooks.

IV.2 Expected Returns after Emergence

In the first approach to measuring bias, we track the behavior of IFC indices right after emergence. These are measured in U.S. dollars, inclusive of dividends. As some of these markets have experienced hyperinflation, it is essential to measure returns either in a common foreign currency or by deflating by the local price index. Both approaches should give similar results in situations where Purchasing Power Parity holds, which is more likely to be the case in hyperinflationary environments.

Hypothesis 1: Expected returns will be higher immediately following emergence than later on.

To test this hypothesis, we adapt the event-study methodology used in the simulation above to the emergence of markets. We construct an equally-weighted index where returns are aligned on the emergence date. The advantage of this "portfolio" approach is that it fully accounts for cross-correlation between events, which is substantial in this case since 7 markets out of 22 emerge on the same date. Figure 11 plots the time-series of the portfolio value. Right after emergence, the portfolio is composed of 22 markets then dwindles to 7 markets after 20 years.

The graph shows that the slope of the line is greater immediately following emergence. This pattern is consistent with the simulation results. The magnitude of the effect is confirmed in Table 3, which considers 36, 48, and 60-month windows after emergence. The portfolio performance is then compared to that over the subsequent window of same length. Assuming independence over time, the standard error of the difference in means is simply obtained from the sum of variances over the two intervals. The table shows that the performance is significantly greater immediately after emergence. The difference is striking: 1.26% (or 15% pa) using 60-month window, 2.56 (or 31% pa) with a 48-month window, and 2.12 (25% pa) with a 36-month window. These numbers are all statistically significant, as indicated by the t-statistics below. Therefore, as suggested by the simulations, we find a significant bias due to recent emergence.

IV.3 Expected Returns around Emergence

Another approach to measuring bias is to recover market information from a completely different source. We take a sample of markets for which equity indices are collected by the International Monetary Fund (IMF) since 1957. These indices are compiled by the local stock exchange and may not be consistent across countries. In addition, they are monthly averages, not end-of-month data. Still, comparisons are appropriate as long as the same IMF series is used before and after emergence. Additional data exist for a total of 7 markets, which are listed in Table 1. The shortest period before emergence is for Peru, for which we have data for two years only.

Hypothesis 2: Expected returns will be higher after than before emergence.

To guard against hyperinflation, we measure returns both in dollars and in real terms (deflated by the CPI as provided by the IMF). As before, returns are aggregated into an equally-weighted portfolio of seven markets aligned on the date of emergence. Figure 12 displays the time series of cumulative returns. The picture clearly indicates a break in trend, with returns after emergence sharply moving upward. As in the previous test, this pattern is consistent with the simulation results reported from our hypothetical market subject to conditioning upon emergence. In fact, Figure 9 is quite similar to Figure 7 which aligns simulation returns on the emergence date.

Formal tests of breaks in expected returns are presented in Table 4. Both real returns and dollar returns strongly indicate that average returns are higher right after emergence. The difference in real returns, for instance, averages 40% annually. Even with a very large standard error of returns of 89% annually, two years of data are sufficient to bring strong rejections of the null. Again, these results strongly suggest that returns are biased upward once a market is considered "emerged."

IV.4 Expected Returns before Emergence

A third approach considers markets that have not yet emerged. Besides its official list of emerged markets, the IFC also collects information on a sample of markets that have the potential to emerge. By the end of 1994, the IFC was watching 24 such markets, for which it provides some annual returns from local stock markets, exchange rates and market capitalization. These markets can be considered "non-emerged". In fact, some of them disappeared for a while, most notably Kuwait, once a $10 billion market.¹¹.

Annual returns were collected for this sample of markets varying from 7 in 1985 to 19 in 1994. To compare their performance with that of established emerging markets, we constructed a value-weighted dollar return index which spanned ten years. ¹². This index was compared to the IFC composite index, which is also value-weighted. To maintain comparability, both indices include only capital appreciation.

Table 5 compares the performance of the two groups of markets, non-emerged and emerged. The table show that the non-emerged group returned an average of 12.5% over these ten years, against 19.1% for the emerged index. This difference confirms, using an entirely different data set, the possibility of biases in the performance of emerged markets. Emerged markets, on average, return 6.6% more than other markets. For comparison purposes, the table also reports the performance of the MSCI World index, a value-weighted index of developed marekts. Over this period the average return was 14.0%, which also falls short of the performance of emerged markets.

In our model, the 6.6% difference can be attributed either to the fact that non-emerged markets truly have low expected returns (low betas in our model) or to the fact that the sample selection process for defining emerging creates survivorship biases. The volatility of the series, unfortunately, is so noisy that the t-test is unable to reject equality of mean returns. Most managers, however, would agree that a difference of about 7% over a decade is economically significant.

V. Conclusion

A very simple model of global markets that allows for differing expected returns provides the basis for simulations of selection of emerging markets. These simulations show an inverse relationship between the recentness of market emergence and the subsequently observed return. The model also shows that recently emerged countries have low covariance with the global market. These results are striking because they fit the empirical observation that emerging markets appear to have high returns and low correlations with other markets. In our baseline model, which assumes no mispricing of emerging markets, some of these high returns are due to survivorship biases.
The findings of these simulations are confirmed by our empirical analysis, which shows that average returns on markets that have just emerged are temporarily high. The history of emerged market provides an overly optimistic picture of investment performance, as we find that other markets do more poorly. Therefore basing investment decisions on the past performance of emerging markets is likely to lead to disappointing results.

A major caveat of this analysis is that it is based upon a stationary model. Economies are never that simple. The global markets have been subject to dramatic changes over the twentieth century, and many nations with bright economic prospects in the 1920s subsequently failed to reward investors for their high expectations. It seems reasonable to condition expected returns in marginal markets on changing political, legal and economic environments. However, it is also important to learn from history. Market contractions and expropriations have occurred in the past, and are likely to occur in the future, even in the absence of a major event such as a world war. If we fail to account for the "losers" as well as the "winners" in the global equity markets, we may be ignoring important information about actual investment risk.

One way to account for losers is to gather additional historical data. Financial economists are accustomed to working with abundant and accurate data, but unfortunately it is strongly conditioned upon survival. For instance, we do not as yet have a CRSP quality data set for Argentina's equity market, even though it has existed since 1926 -- at which time Argentina was one of the world's major economies.

In all likelihood, we will be forced to rely upon poorer quality data to draw inferences about markets in the state of crisis. Much "financial archaelogy" remains to be done. We have begun to collect data on global markets that have emerged and submerged since 1929. A close look at this data indicates that it is based upon poorly documented indexing methodologies. It also typically lacks dividend information. Nonetheless, the results of our simulations suggest that it is better to look at lesser-quality data than to ignore it. It is precisely the hard-to-get data that tells us what happens when a market submerges.

References

Bekaert, Geert and Campbell Harvey, 1995a, "Time-Varying World Market Integration," Journal of Finance 50, 403-444.

Bekaert, Geert and Campbell Harvey, 1995b, "Emerging Equity Market Volatility," Working Paper, Stanford University.

Brown, Stephen J., William N. Goetzmann, Roger G. Ibbotson and Stephen A. Ross, 1992, "Survivorship Bias in Performance Studies," Review of Financial Studies 5 , 553-380.

Brown, Stephen J., William N. Goetzmann and Stephen A. Ross, 1995, "Survival," Journal of Finance 50, 853-873.

Ferson, Wayne and Campbell Harvey, 1991, "The Risk and Predictability of International Equity Returns"

Goetzmann, William and Philippe Jorion, 1995, "A Longer Look at Dividend Yields," Journal of Business, 483-508.

Harvey, Campbell, 1991, "The World Price of Covariance Risk," Journal of Finance 46, 111-157.

Harvey, Campbell, 1995, "Predictable Risk and Returns in Emerging Markets," Review of Financial Studies 8, 773-816.

International Finance Corporation, 1995, The IFC Indexes: Methodology, Definitions, and Practices IFC: Washington, DC.

Jorion, Philippe, 1985, "International Portfolio Diversification with Estimation Risk," Journal of Business, 259-278.

Park, Keith and Antoine Van Agtmael eds., 1993, The World's Emerging Stock Markets, Chicago: Probus.

O'Conner, Selina and David Smith, 1992, The G.T. Guide to World Equity Markets, London: Euromoney Publications.

Notes

* Thanks N. Prabhala for helpful conversations on this topic. Due to the preliminary nature of this research, results are likely to change.

1.See for example, Bekaert and Harvey (1995a and 1995b), Ferson and Harvey (1993), Harvey (1991), and Harvey (1995).

2. See Brown, Goetzmann and Ross (1995).

3. Bekaert and Harvey (1995), presents an argument that this local factor (or factors) may be priced in the absence of full market integration. That is, with barriers to outside investment and diversification, standard asset pricing may not apply. This interesting issue is not addressed in our stylized model. We assume that the returns are those that accrue to outside investors who are able to diversify idiosyncratic risk sufficiently for the standard models to apply.

4. This inverse relationship between beta and last emergence is also true for R-square as well. Early emergers have an R-square of 0.7 generated by a regression of the time-series of returns on the global market index. The most recent emerging markets in our simulation have an R-square of under 0.5. This also appears consistent with the empirical results in Bekaert and Harvey (1995).

5. This survival problem might be mitigated were we able to collect return data on the submerged series--not an easy matter, when investors have lost interest, and/or the market has closed.

6.The BGR analysis suggests an equivalence between distance from the bound and the residual variance of the series, and the use of the appraisal ratio, i.e. alpha devided by residual risk, is shown to minimize certain forms of survival bias.

7. Harvey (1991) shows this assumption may be violated due to time-variation in global market premia.

8. The results of a t-test for differences in means between the average annual return of the market up-to and including last emergence vs. average annual returns following last emergence yield a t of 179 on 35606 degrees of freedom. The mean before emergence is -.1077, and the mean after emergence is .1339. Obiviously, due to the large sample size of the simulation we are able to reject equality with a high degree of confidence.

9. Jorion (1985) pointed out that the practical application of mean-variance optimization to international diversification is seriously hampered by estimation risk.

10.Countries with price indices compiled by the League of Nations include: Canada, U.S., Chile, Colombia, Mexico, Uruguay, Venezuela, Japan, Germany, Belgium, Denmark, Spain, Finland, France, Hungary, Ireland, Norway, Netherlands, Portugal, Romania, UK, Sweden, Switzerland, Czechoslovakia, Australia, New Zealand, Greece, Italy, and Peru.

11.Countries watched by the IFC in 1994 include Bangladesh, Barbados, Botswana, Costa Rica, Ivory Coast, Cyprus, Ecuador, Egypt, Ghana, Honduras, Iran, Jamaica, Kenya, Kuwait, Mauritius, Morocco, Namibia, Oman, Panama, South Africa, Swaziland, Trinidad, Tunisia, and Uruguay. Of those, Costa Rica and Honduras have no stock price index, the series for Uruguay stops in 1991, and the series for Kuwait was interrupted from 1990 to 1993 because of the Gulf War.

12.In the computation of the value-weighted index, we omit South Africa, because its market value would dwarf all others. As of December 1994, the market capitalization was $226 billion, while that of the next largest market is $10 billion.