Author: Eli Bartov
Leonard M. Stern School of Business at New York University
Discussant is Bob
Jensen
Trinity University
Conference on International Accounting Related
Issues
Hosted by Haim
Falk
School of Business
Rutgers University at Camden
Camden,
NJ 08102
May 31-June 1, 1997
I (Bob Jensen) would like to begin by thanking Haim Falk for inviting me to discuss Professor Bartov's paper at this conference. It is an honor to be in the company of such noted accounting and finance researchers.
My discussion will focus upon three major categories:
Overview
of Professor Bartov's Paper
Comments
Upon Model and Empirical Outcomes
SQL vs
Accounting Measurement Rules of the Future
Overview of
Professor Bartov's PaperThis paper follows up earlier research on the association between stock prices and the 1981 implementation of SFAS No. 52 in place of the controversial SFAS No. 8 for foreign currency translation. Under SFAS No. 52, firms who elect the dollar as the functional currency for foreign subsidiary financial reporting, the SFAS No. 8 temporal method is still in effect. However, for the majority of firms who select the foreign currency as the functional reporting currency, SFAS 52 permits the use of the current method and deferral of gains and losses in a special equity account. Details comparing the temporal versus current methods are given in a book called International Accounting by Choi and Mueller. The principle impact of SFAS No. 52 is to allow firms to value assets and liabilities at current currency exchange rates and to defer the impacts of rate fluctuations on the income statement by using a special equity account for such purposes (rather than retained earnings).
Under the temporal method, monetary items are translated at current rates whereas nonmonetary items are translated at historical rates. On Page 4, Professor Bartov notes that:
If, however, nonmonetary assets and liabilities provide a (natural) hedge for each other, the reported translation gain or loss of the temporal method will be misleading, as it fails to account for the fact that translation gains (losses) on liabilities are hedged (i.e., offset) by correlated translation losses (gains) on nonmonetary assets.
SFAS No. 8 was highly criticized by multinational corporations when it was introduced in 1976. The main criticism was the fluctuation in earnings caused by the SFAS No. 8 temporal method. In 1981, SFAS No. 52 was a much more popular change in the measurement rules, although firms with the same industries did not all abandon the temporal method. This has led to some confusion and suspected inconsistencies with respect to choice of functional reporting currencies.
Professor Bartov does an excellent job of reviewing the earlier research and pointing out the need for a more extensive inquiry into whether the implementation of SFAS No. 52 had a marked impact upon stock prices. He also points out that his is not an "events study" surrounding the initial impact of a firm's SFAS No. 52 adoption and stock prices over a short time period (e.g., a few days). Rather, his is an "association study" of the longer range association of stock prices and currency translation choices over a period from 1976 through 1990.
In relation to earlier studies using the CAR regression model, Professor Bartov is careful to avoid some of the weaknesses of earlier studies, notably the Collins and Salatka paper appearing in a 1993 issue of Contemporary Accounting Research.
- Foreign currency translation adjustments are separated from earnings before being used as explanatory variables in the model.
- A longer term association study is conducted as opposed to an events study over a time frame of, say, two days following each earnings announcement.
- Levels of adjustments rather than first differences are used.
- Earnings surprises are included the the Equation 2 model on Page 12.
- Suspected nonlinearity problems in Equations 1 and 2 are dealt with by using rank regression to avoid having to specify a nonlinear model. Explanatory variables are replaced by their ranks and scaled by the number of observations per year (see Page 14).
- Correlation coefficients are compared in Table 4 to test for multicollinearity.
- A dummy variable is added to Equation 2 that allows the foreign currency adjustment variable (RCURADJ1) to vary for firms having a net foreign currency asset position versus a liability position. See Page 18.
Comments Upon
Model and Empirical Outcomes
My Concerns About the Use of the CAR Regression
DatabaseI feel a bit like Charlie Brown explaining how best to kick a football with Lucy as a placeholder. The Lucy in this case is Professor Bartov who may well pull the football away from my attempts to kick him about in this discussion. He is much more experienced and learned in capital markets research using the CAR model (see Equation 2 on Page 8 of his paper).
The Capital Asset Pricing Model (CAPM) underlies the the CAR regression model in the CRSP NYSE/AMEX's daily beta excess returns for the firms in the sample. I have little experience using the CRSP database and a somewhat negative suspicion of studies based upon beta excess returns of a single index model subject to severe structural modelling error. However, since the limitations of beta excess returns are so well documented in the literature, I will not repeat those serious doubts and concerns here. Suffice it to say that there is a large body of important empirical research accumulated using the CAR regression database such that Professor Bartov's paper can be compared in relation to earlier studies. Little can be gained today by repeating the debate over CAR models in general.
My Concerns About the Use of Ratios as Independent
VariablesRatios create special problems in multivariate analysis. Professor Bartov uses ratios because the linear relation is suspect and the form of the nonlinear structure is not practical to estimate. Earlier writers attempt to overcome this problem by using rank regression in place of interval-scale regression. Variates having a nonlinear relation can have a linear rank relationship. However, ranks themselves are dependent upon the sample size. Professor Bartov uses the approach of replacing explanatory variables by their ranks, the ranks themselves are dependent upon the sample size. For instance, a rank of 1 in a sample of N=10 is not the same as when the sample is N=1000.
I have some concerns about the use of ratios that are not discussed in the paper. First there is the problem of truncation. The normal distribution assumes that variates are not bounded, whereas transformations used by Professor Bartov are bounded by zero and one. Secondly, there is danger in comparing outcomes from different sample sizes. The ratio of 1/10 yields a rank of 0.100 when the rank is 1 among N=10 sampled items. However, it becomes 0.001 when N=1000. When N=10 the only possible ratio outcomes are 1/10, 2/10, 3/10, 4/10, 5/10, 6/10, 7/10, 8/10, 9/10, and 10/10 assuming no ties. There are only 10 possible outcomes when N=10. However, when N=1000 there are 1000 possible ranks beginning with 1/1000 and ending with 1000/1000. Suppose the rank 0.100 in the N=10 sample is compared with the rank 0.001 in the N=1000 sample. It is not even possible to obtain a rank of 0.001 in the small N=10 sample. The scaling by sample size in this case hardly leaves the two numbers 0.100 and 0.001 comparable. Before and after the ratio scaling, both outcomes are at the top rank in their respective samples. Yet there is a large difference between 0.100 and 0.001.
Therefore, it is not clear to me what is gained by replacing ranks with their ratios. It seems that the use of ratios clouds rather than helps when evaluating ranks. If the ratios are later compared it may be analogous to using sows' ears for silk purses. We cannot make interval or ratio scaling of the ratios of ranks. For example, the distance between a rank of 0.100 and 0.001 is a meaningless and possibly misleading value since both outcomes have Rank 1 in their respective samples. The 0.100 and 0.001 scores are both the top ranks in their respective samples.
More importantly, one cannot assume interval or ratio scaling of the ratios of ranks. For example, the distance between a rank of 0.100 and 0.001 is a meaningless and possibly misleading value. By whatever value, these outcomes are still the top ranks in their respective samples.
My Concerns About Statistical Inference Testing of
Large SamplesI am always suspicious of statements that outcomes are "statistically significant" when inference tests are run on very large samples. Outcomes that are insignificant from a decision standpoint are often "statistically significant" in large samples. For example, in a huge sample of third grade students, suppose girls might exceed boys by 0.00000067981 in IQ testing. Even if that rejection of a null hypothesis (that their is zero difference in IQs between boys and girls) is statistically significant, this rejection is totally irrelevant from a decision making standpoint.
If the null hypothesis is that the mean of a population is a hypothesized value, the probability of accepting the hypothesis based upon the difference between the sample mean and the hypothesized mean varies considerably among small samples but not between large samples. The typical Operating Characteristic (OC) curves depicting the probabilities of accepting such a hypothesis are shown below:
In the above graphs, note that the beta probability of acceptance of a null hypothesis approaches zero as sample sizes become larger and larger. More importantly, note that if there is any divergence between sample and hypothesized means, in extremely large samples the null hypothesis will always be rejected rather than accepted since the probability of acceptance on an OC curve approaches zero asymptotically. For large samples, of say 2,500 or more items, the OC curves almost coincide with the ordinate in the above graph making acceptance of the null hypothesis virtually impossible.
Professor Bartov uses the Wilcoxan Signed-Rank test which becomes asymptotically normal for sample sizes of N>25. In Table 6, the null hypothesis is not rejected for that subset of 378 firms who used the US dollar as the reporting functional currency.
However, in Table 6 Professor Bartov does reject the null hypothesis of no association between SFAS No. 52 adjustment and stock prices for sample sizes of 2,071 for firms who use a foreign functional reporting currency (see Table 2). The question becomes whether this rejection of the null hypothesis is due to substantive differences in addition to the huge sample size. There are so many transformations of the variables (ranks and then ratios of ranks based upon sample sizes) in a complex CAR rank regression model, that the magnitudes of the regression coefficients are very difficult to evaluate in terms of "substantive" tests. I will leave it up to Professor Bartov to comment upon the substantive outcomes apart from the statistical inference tests. In other words, do we have a phenomenon analogous to an observed 0.00000067981 difference in IQs between boys and girls that is "statistically significant" but not "substantively significant?" Or do we have an outcome that is both statistically and substantively significant?
To put large sample estimation in another context, suppose tests are run for confidence interval estimation. Using Excel, I derived a sequence of confidence interval estimates as the sample increases from N=5 to N=10,000 cases. Note how quickly the confidence interval shrinks to where it nearly evaporates.
Statistical inference is often more misleading than helpful for very large samples. When samples are large it is nearly impossible not to reject null hypotheses and/or obtain useful confidence intervals.
SQL vs Accounting
Measurement Rules of the FutureI predict that events and association studies of accounting rule impacts will become increasingly and, in many instances, impossible in the future. Networking and database technologies will make it impossible to isolate the impact of one nation's change in an accounting reporting rule. For example, at the moment in the 1996 Financial Statements of Microsoft Corporation, it is possible to obtain the 1996 Income Statements under the reporting rules of six different nations.
|
|
For over a decade, a number of writers have been predicting that corporations will make databases available such that investors may compile customized financial reports based upon alternative accounting measurements. In 1990, Professor Beaver at Stanford University sent me a working paper entitled "Boundaries of Financial Reporting and Future Events." In that paper on Page 16, he states:
While such questions are not easily resolved, they are considerably different from the issues the FASB now devotes major time and effort ... in attempting to resolve, such as trying to determine which set of accounting methods produces the "best" measure of net income. In the database setting, there would be less concern with how to calculate the "bottom line" and more concern with the nature of disclosure presented from which users can make their own calculations.
What nobody predicted was the dramatic changes in technologies that would make such database reporting so quickly feasible, practical, effective, and efficient to a point where the barriers of attitude and skills are more serious than hardware and software barriers. The major technology changes include:
- Explosion of Corporate Web Sites
- Virtually all publicly-traded corporations now have web sites in anticipation of an explosion of customers, investors, and prospective employees on the Internet. Click here for the Yahoo search site for companies.
- Explosion of Investors and Analysts Using the Internet
- Following the immense popularity of EDGAR from the Securities and Exchange Commission and the rise in popularity of financial markets news and services on the Internet, the number of users of these sites has exploded exponentially with no end in site. Furthermore, the explosion encompasses investors from all over the world who no longer tolerate delays in accessing financial information around the globe. For links to such services, see my Document 3 at http://www.trinity.edu/~rjensen.
- Decline in Cost of Data Storage on Web Servers
- Costs of storage of massive amounts of data on web servers have declined dramatically. Storage costs are so inexpensive that it is now possible to accompany text and data with graphics and audio in networking, including "real" audio that does not require downloading of entire files. Even more important, is the rise of software such as that from Bell Labs that will synthesize web audio from text without having to store any audio files. What this means is that accounting data can be stored in databases along with accompanying text and audio instructions about how to derive customized accounting reports under alternative measurement rules and standards of other nations.
- Evolution of Relational Databases
- The single most important advance in database technology over the past two decades has been the rise of what is known as a "relational database." In my Document 12a Technology Glossary, I define a relational database as follows:
A database system that stores data in two-dimensional data tables at the same time such that the program can work with two tables at the same time. It is "relational" if one table defines the relation between entries in rows (data records) and columns (fields). Not all database software claiming to be relational meet the "true" relational database mathematical theory developed by Edgar Cobb in 1970. For example, dBASE and FoxPro can link two databases through a common field but are not true relational database programs.
- Emergence of SQL
- The most important extension of relational database technology is what is known as Structured Query Language (SQL). The FOLDEC Free On-Line Dictionary of Computing at http://theory.doc.ic.ac.uk/foldoc/ defines SQL as follows:
A language which provides a user interface to relational database management systems, developed by IBM in the 1970s for use in System R. SQL is the de facto standard, as well as being an ISO and ANSI standard. It is often embedded in other programming languages. The first SQL standard, in 1986, provided basic language constructs for defining and manipulating tables of data; a revision in 1989 added language extensions for referential integrity and generalized integrity constraints. Another revision in 1992 provided facilities for schema manipulation and data administration, as well as substantial enhancements for data definition and data manipulation. Development is currently underway to enhance SQL into a compositionally complete language for the definition and management of persistent, complex objects. This includes: generalization and specialization hierarchies, multiple inheritance, user defined data types, triggers and assertions, support for knowledge based systems, recursive query expressions, and additional data administration tools. It also includes the specification of abstract data types (ADTs), object identifiers, methods, inheritance, polymorphism, encapsulation, and all of the other facilities normally associated with object data management.
The implications of this rise in networking, hardware, and software technology for financial reporting are mind boggling. In 1981, US multinationals spent considerable time and money both in persuading the FASB to override FASB No. 8 with FASB No. 52. In the next decade, there will not be a need for such concern since investors will be able to choose the foreign currency translation method they prefer when customizing the financial statements of a multinational corporation from that company's web server.
Will database reporting eliminate the need for the
FASB, IASC, and other standard setting bodies?The standard setting bodies will be remain, but the focus of their attention will shift from measurement rules to database disclosure rules. For example, some disclosures are relatively costless. Foreign currency disclosures fall into this category, because both historical exchange rates and current exchange rates are relatively costless to obtain and store based upon daily currency exchange market data. However, data that is costly to obtain such as replacement and exit value estimates of operating assets may require guidance and perhaps rules regarding the amount and quality of the data to be provided in web servers.
What are the implications for international
harmonization?In the past, international harmonization has focused largely upon measurement rules regarding what gets booked as assets, liabilities, expense, and income. In the future, harmonization will focus almost exclusively upon what gets disclosed since investors around the world will be able to use relational databases, SQL, and other software for customizing data derived from harmonized databases.
What are the main concerns of corporations regarding
database reporting?The reluctance of corporations to make internal databases available to the public is the main barrier to the rise of database reporting of investment data. Managers are torn between conflicting goals of drawing in more investment capital at lower costs versus fears that additional disclosures will aid competitors and turn away some investors who learn more about companies than is good for those companies who air too much dirty linen. Equally important is the fear that the public will not be able to differentiate between hard numbers such as cash in bank accounts versus soft numbers such as claims filed by or against a company in civil courts. Amidst the rising flood of information, database reporting adds more pools and flow than many investors are able to manage. Sophisticated investors with lots of computer expertise may gain comparative advantages over those investors who have never heard of a 56K modem. On top of this is the very real concern over the security and integrity of the databases themselves and the risk of errors and fraud by high-tech trouble makers that have done such things as create imposter web sites for President Clinton and the Central Intelligence Agency.
What are the main concerns of investors regarding
database reporting?Investors have some of the same concerns as corporate managers. How will investors cope with the rising flood of information? How will technology disadvantaged investors avoid being exploited by their high-tech counterparts? To this we add special concerns of investors regarding any technologies that increase the risks of error and fraud on the part of corporate managers themselves.
Many of the above questions are taken up in greater detail in a forthcoming paper by Jensen and Sandlin that is entitled "The Paradigm Shift in Technology: Financial Reporting Will Never Be the Same." This paper is forthcoming in the next issue of Research on Accounting Ethics in 1997 by JAI Press Inc.
My main message for the audience of this conference is that many of the papers focused upon in this conference are probably a dying breed due to the monumental impacts that technology will have upon reporting of accounting and other data. For example, if both historical and current exchange rates of assets and liabilities were reported in databases, it would be pointless to try to examine the CAR impacts of translation rules since investors would have access to outcomes under alternative rules.
Once again, I want to thank Haim Falk for inviting me to this conference and for giving me a chance to learn from the carefully crafted paper of his friend Eli Bartov. I suspect that some of my more direct concerns regarding CAR modelling, ratio transformations of ranks, and sample sizes will be refuted by Professor Bartov since he is more into CAR modelling than me. I think, however, that my prognosis for the future will greatly change the nature of empirical research in capital markets. I remind you once again that Microsoft Corporation is already reporting income statements under the differing reporting rules of six nations. With just a click of the mouse, investors can compare the results of different reporting rules on the 1996 income statements of Microsoft Corporation. This is the beginning of the end for heated debates over requiring a single standard such as SFAS No. 8 or SFAS No. 52.
