International Journal of Forecasting 31 (2015) 392–398
Contents lists available at ScienceDirect
International Journal of Forecasting journal homepage: www.elsevier.com/locate/ijforecast
Editorial
International Financial Forecasting: Global Economic Linkages and Corporate Earnings The International Journal of Forecasting (IJF ) has a long history of publishing studies on the efficiency of capital markets and earnings forecasting. Brown (1993) linked earnings forecasting to capital market efficiency, and later edited a Thomson Reuters Annotated Bibliography that summarized over 600 studies of earnings forecasting (Brown, 2008). Ramnath, Rock, and Shane (2008) updated and addressed many of the issues in earnings forecasting that were suggested by Brown. More recently, Guerard, Rachev, and Shao (2013) and Guerard, Markowitz, and Xu (2013) published research reporting that global earnings forecasting continues to drive global stock returns. In 1998, Brown and Guerard co-edited a special issue of the IJF on ‘‘International Financial Forecasting’’ that addressed foreign exchange forecasting and earnings forecasting (Brown & Guerard, 1998). The issue featured articles by Modigliani and Askari on the increasing volatility of exchange rates during the first twenty-five years of flexible exchange rates, and by Enders and Falk on the lack of forecasting efficiency from the purchasing power parity theory of exchange rates. Conroy, Harris, and Park reported that fundamental factors continued to drive Japanese equity prices, along with management forecasts of earnings. Good news stocks were defined to be those where actual earnings exceeded analysts’ forecasts by more than 10%. Bad news stocks were defined to be when actual earnings were more than 10% below analysts’ forecasts. Good news earnings surprise Japanese stocks slightly outperformed bad news earnings surprise Japanese stocks. Guerard, Blin, and Bender reported that I/B/E/S forecasts of earnings enhanced returns for a Japanese market-neutral portfolio relative to using a stock selection model based on earnings and book values, whereas the I/B/E/S earnings forecasts were necessary for a U.S. market-neutral model to outperform; that is, the value-only U.S. market-neutral stock model did not outperform the LIBOR rate, its benchmark, but the addition of the earnings forecast variable created a profitable strategy. This issue continues in much the same spirit as the 1998 special issue. Ziemba and Lleo, in ‘‘Some Historical
Perspectives on the Bond-Stock Earnings Yield Model for Crash Prediction around the World’’, study stock market crashes using a simple model with only a single variable, the difference between the stock and bond rates of return (BSDM). The idea is that stocks and bonds compete for investment dollars, with stocks being favored when interest rates are low and bonds being favored when interest rates are high. When the long bond yield minus the earnings yield (the reciprocal of the price earnings ratio) is very large, there is a high chance of a stock market crash. The model explains the October 1987 crash. Ziemba and Lleo demonstrate that there is extreme danger in the stock market at such times because the 30-year government bond yields are much higher than the usual stock market yields, measured by the reciprocal of the previous year’s reported price earnings ratio. These high interest rates invariably lead to stock market crashes. Here, the danger indicator moved across a statistical 95% confidence line in April. The market ignored this signal and eventually crashed in October 1987. There was a similar signal in the US S&P500 around April 1999 that was ignored by most investors, followed by a crash that began in August 2000 and a weak stock market in 2001/02. Ziemba and Schwartz (1991) found that there were twenty 10-plus crashes during the forty-year out-ofsample forecast period, 1949–1989. The Ziemba and Lleo BSDM predicted 12 out of 12 market crashes, a splendid prediction record. Not all crashes had the measure in the danger zone, but whenever it was there was a crash, with no misses. An additional eight crashes occurred for other reasons. However, long-run mean reversion suggests that the longer the bull run is and the more over-priced the measure becomes, the longer and deeper the decline will probably be. One can then use the measure as part of an econometric system for estimating future scenarios. Levanon, Manini, Ozyildirim, and Schaitkin, in ‘‘Using Financial Indicators to Predict Turning Points in the Business Cycle: The Case of the Leading Economic Index for the United States’’, review financial, monetary and credit market indicators in the global context, from the
http://dx.doi.org/10.1016/j.ijforecast.2015.01.001 0169-2070/ © 2015 Published by Elsevier B.V. on behalf of International Institute of Forecasters.
Editorial / International Journal of Forecasting 31 (2015) 392–398
perspective of their relationships with the general business cycle of the U.S. economy. They document which of these financial indicators are useful in predicting recessions and recoveries (i.e., business cycle turning points), and argue that aggregating our selected indicators into a composite index offers advantages over relying on them individually. Given the nature of most of the indicators selected, they call this composite index the leading credit index (LCI). The advantages of the LCI are due to the ability of the simple, easy-to-calculate and transparent methodology of the composite index approach to generate reliable and smooth estimates of an unobserved business cycle variable. Their proposed index is the principal component of six selected indicators, and thus, it aggregates various different types of quantitative and qualitative survey indicators, which are all related to the availability and cost of credit, and economic agents’ willingness to borrow or lend. They argue that this new index captures important channels through which the financial sector can impact the real economy. They also show that the suitability of one of the financial components of the current LEI as a leading indicator, namely the real money supply, as measured by M2, has declined in recent decades as a result of changes that have occurred in the U.S. economy. They argue that their leading credit index is an appropriate replacement for the money supply component of the LEI. They show that the forecasting performance of the leading index can be improved if their new index of financial conditions is used as a component of the LEI, replacing the money supply measure that is currently used as a component. The contribution of the new LCI to forecasting during the great recession is noteworthy. It is also important to note that the forecasting tests are constructed with the realtime performance of the LEI in mind, and that they are not used in the indicator selection process. Considering their overall findings on the new leading index of financial indicators proposed in this paper, the authors believe that the LEI can provide real forecasting improvements in forecasting both growth and turning points. The real time out-of-sample forecasting performance of the new LEI (that replaces the real M2 component with this new index) supports this conclusion. The changes in the behavior and usefulness of real money supply (i.e., it having ceased to be a useful leading indicator) for monitoring and predicting the economic cycle, and the emergence of new and more useful financial indicators, are results of structural changes in the U.S. economy and finance markets that have taken place in a global context over the last two to three decades. In ‘‘A Further Analysis of The Conference Board’s New Leading Economic Index’’, Lahiri and Yang assess the marginal impact of the latest change in the composition of the Leading Economic Index, as proposed in the accompanying paper by Levanon, Manini, Ozyildirim, Schaitkin, and Tanchua (2015), from the standpoint of probabilistic business cycle predictions and using a battery of diagnostic tools. By recognizing the binary nature of the target variable, these evaluation metrics supplement the quadratic probability score that was used extensively by Levanon et al. (2015). Their analysis reveals additional insights into aspects of the forecast accuracy that are not measured directly by the QPS. They show that the better performance
393
of the new LEI is attributable mainly to its greater discriminative capacity for distinguishing recessions from expansions. By examining conditional probability functions and ROC curves, the relative superiority of TCB’s new LEI is established firmly. Levich and Poti, in ‘‘Predictability and ‘Good Deals’ in Currency Markets’’, estimate predictive models for the excess return on the ith currency of the general ARMA(p, q) model, where p denotes the autoregressive (AR) lag order and q denotes the order of the moving average (MA) term: ri,t = b0 + b1 ri,t −1 + · · · + bp ri,t −p + c1 ui,t −1 + · · ·
+ cq ui,t −q + ut .
(1)
The above predictive regression is intended as a reducedform representation of the DGP of the (excess-) return on the ith currency over the estimation window from t0 to t1 , with t0 ≤ t ≤ t1 , where t0 and t1 ∈ {1, 2, . . . , T } and T denotes the end of our overall sample period. In conducting the AIC-based model selection, they allow for AR and MA terms p and q of up to the fifth order. They fit versions of Eq. (1) to both raw returns and returns adjusted by the interest rate differential (i.e., the differential between the interest rate on deposits denominated in the currency and the US Dollar funding cost). In this paper, they assess the statistical and economic significance of predictability in currency returns over the period 1972–2012 by analysing the frequency and magnitude of violations of a boundary that is consistent with rational pricing and the absence of ‘‘good deals’’. While data on a suitable risk-free asset proxy are not available throughout the sample period, they find that adjusting returns for the interest differential has virtually no impact on the estimated predictability. Daily and monthly predictability and excesspredictability estimates are reported. In the daily case, Levich and Poti set p = 5 and q = 0 in the predictive ARMA(p, q) because, for most sample periods and currencies, these are the parameter values selected by the AIC and this seems a natural choice at the daily frequency. They report the AR and MA orders as well as the coefficients of determination and the related excesspredictability measures of the selected ARMA (p, q) models of monthly returns. The reported excess-predictability measures are sizable in many cases, suggesting the attainability of excess-SRs that are even larger than 100 per annum, especially for daily returns in the initial part of the sample period. The differences between the initial and subsequent parts of the sample period appear to be less pronounced in the monthly than in the daily case. Levich and Poti find that, even under a relatively wide bound on the relative risk-aversion, predictability often violates the attendant theoretically-motivated upper bound. This happens in as many as 10% to 15% of the rolling estimation windows, though typically only briefly. This evidence implies the availability of ‘‘good deals’’, and thus the violation of the EMH under a broad class of asset pricing models, with conservatively high values of the marginal investor’s RRA and realistic levels of transaction costs. Enders and Pascalau, in ‘‘Pretesting for Multi-StepAhead Forecasts with STAR Models’’, focus their attention on STAR models because they are particularly well-suited to the characterization of important macroeconomic
394
Editorial / International Journal of Forecasting 31 (2015) 392–398
variables such as the unemployment rate, the growth rate of industrial production, and exchange rates. Moreover, the multi-step-ahead forecasts from the STAR-type models that they consider ease the specification problem greatly in that the form of the model does not change at different forecasting horizons. The data consist of monthly observations over the period January 1975 to December 2013. They have included only the period of flexible exchange rates. However, for most European countries, the sample ends in the last quarter of 1998, when the Euro replaced the national currencies of the countries that joined the single monetary system. For completeness, they employ three different measures of each country’s real exchange rate in relation to the U.S. dollar. The first measure is the bilateral real exchange rate, constructed as the natural log of the product of the number of foreign currency units per U.S. dollar and the U.S. Producer (Wholesale) Price Index (PPI/WPI) divided by each nation’s WPI/PPI. The second version constructs bilateral real exchange rates using Consumer Price Indices, also in log form. The third exchange rate measure is the natural log of the real effective exchange rate (REER), as constructed by the IMF. Enders and Pascalau report findings from applying their test to each of the real exchange rate series at the one-, four-, eight-, and twelve-period forecasting horizons. Note that their test at the one-period horizon is equivalent to Teräsvirta’s (2006) in-sample test. The appropriate order of integration of the real exchange rates is debatable. However, they follow traditional forecasting procedures and first difference all of the real rates in order to ensure stationarity. A majority of the CPI-based RERs display linear adjustment at all forecasting horizons, while a majority of the PPI-based RERs and REERs display some type of nonlinear adjustment, at least for one of the forecasting horizons. Across all three exchange rate types, most of the instances where the null of linearity is not rejected occur in countries that are part of the Euro zone. It is well known that multi-step-ahead forecasting using nonlinear models entails ‘‘potentially serious practical challenges’’. Not only are such forecasts computationally difficult, there are also, as Teräsvirta (2006) showed, alternative choices that can lead a researcher astray when using a nonlinear model to make multi-step-ahead forecasts. To circumvent some of these problems, Enders and Pascalau propose a simple pretest to determine whether it is better to forecast using a linear or a nonlinear model. The advantage of their pretest is that one might be able to avoid estimating and fitting a nonlinear model when the relationship between the current and future values of a series is sufficiently linear. The (non)linearity test of this paper extends that of Teräsvirta (1994) in order to consider nonlinearity at various forecasting horizons. Depending on the order of the expansion, one can capture both forms of a STAR process (ESTAR and LSTAR). The size and power of the test are quite reasonable. The empirical section considers several types of real exchange rates from a sample of OECD countries. They assess the properties of the tests using out-of-sample forecasts at one, four, eight, and twelve forecasting steps. They use both the random walk model and linear autoregressive models
as benchmarks for evaluating the test. Among their main findings, they show that when the proposed test strongly rejects the null of linearity, (i) a direct forecasting method outperforms a linear iterated one, (ii) the bootstrap predictor of a nonlinear model (ESTAR or LSTAR) outperforms the direct method, and (iii) the best nonlinear model outperforms the best linear model in terms of having lower mean absolute percentage errors. However, when the null is either not rejected or only mildly rejected, the best linear method yields forecasts that are slightly superior to those generated from a nonlinear model. These results confirm prior evidence and provide support for their proposed methodology. Finally, a set of robustness checks confirms that the forecasts from a nonlinear (or linear) model display lower RMSEs and MAPEs than those of a random walk model. Ye, Ashley, and Guerard, in ‘‘Comparing the Effectiveness of Traditional vs. Mechanized Identification Methods in Post-Sample Forecasting for a Macroeconomic Granger Causality Analysis’’, investigate the impacts of different model identification methods on post-sample forecasting. In particular, they identify forecasting models using two different approaches: a traditional, partially judgmental method and the mechanized Autometrics method. They then compare the effectiveness of the different models in the specific context of the post-sample forecasting used in completing a relatively large scale macroeconomic Granger causality analysis. The model parameters are first estimated on the sample running from 1960:2 to 1992:12 and used to produce a forecast for each endogenous variable at date 1993:1, then re-estimated each month and new forecasts produced. The corresponding (rolling) one-step-ahead forecast errors are then used to compute the post-sample mean squared forecast error (MSFE) for each of the four endogenous variables, using both the unrestricted and restricted models for that variable. They also construct naïve benchmark forecasts (intercept-only models, corresponding to a constant growth rate or change) for each of the four endogenous variables, then compare the post-sample MSFEs from these naïve forecasting models to those from both the restricted and unrestricted models. In addition to these forecasting results over the entire post-sample period (i.e., from 1993:1 to 2013:5), they also compute post-sample MSFE results for two subsets of this period: a ‘‘pre-crisis’’ period (1993:1 to 2007:12) and a ‘‘crisis-plus-aftermath’’ period (2008:1 to 2013:5). Regardless of which model identification approach is used, they find that both the restricted and unrestricted models are able to produce more accurate forecasts than the naïve model in most cases, and that the forecasts for the crisis-plus-aftermath period (2008:1–2013:5) are generally less accurate than those for the pre-crisis period (1993:1–2007:12). Notably, the post-sample MSFE results from the models based on the Autometrics specification algorithm are always larger than those from the partially judgmental model specification approach. While it is not clear whether these differences are statistically significant, the uniformity of the results strongly suggests that the ‘‘informed common sense’’ utilized in the partially judgmental model specification method yields better
Editorial / International Journal of Forecasting 31 (2015) 392–398
models, in terms of post-sample forecasting abilities, than does the current state-of-the-art in mechanical model specification methodology. Ye et al. find that the post-sample forecasting ability of the models identified by the traditional method is generally superior to that of the models identified by the mechanized method. In terms of specific Granger causality testing results, the traditional, partially judgmental model identification method yields statistically significant post-sample evidence of Granger causality running from consumption to income, from income to the inflation rate, and from consumption to changes in the unemployment rate. In contrast, a completely analogous analysis using forecasting models identified using the mechanized Autometrics method finds only weak evidence of consumption Granger-causing changes in unemployment; the difference in this set of Granger causality results is a consequence of the mechanically-produced model specifications being less able to forecast post-sample. Overall, they find that the choice of model identification method does indeed have a notable impact on both post-sample forecasting and Granger-causality testing results. In particular – for better or worse – a bit of experienced human judgment still yields better forecasting models than does the best currently-available mechanical method, at least for this particular data set. Brown and Zhou, in ‘‘Interactions between Analysts’ and Managers’ Earnings Forecasts’’, examine the interactions between analysts’ and managers’ earnings forecasts with respect to financial statement information (FSI) compared with the information underlying stock returns (SRI). They show that (i) managers’ comparative advantage over analysts is greater for incorporating SRI into their forecasts than FSI; and (ii) after observing management forecasts, analysts improve their earnings forecasts primarily by a better utilization of SRI rather than FSI. Specifically, they investigate the comparative efficiencies of analysts and managers at incorporating FSI versus SRI into their forecasts; and the relative importance of FSI versus SRI in explaining how analysts improve their forecasts subsequent to managers’ forecasts. Brown and Zhou contend that, in incorporating SRI into forecasts, managers and analysts observe certain events and evaluate the impacts of these events on future earnings. They also contend that capital markets incorporate these events into stock returns, and therefore they use stock returns as a proxy for these events. Managers’ comparative advantages over analysts may or may not be greater with respect to incorporating SRI into their forecasts rather than FSI. On the one hand, managers’ close involvement in their firms’ operations and decision making gives them an advantage in distinguishing the information in past returns that is relevant for the prediction of future earnings. FSI is distributed and perused widely by analysts, so managers may not have any information advantage over analysts with respect to FSI relative to SRI. On the other hand, managers have more control over FSI than over stock returns, so they may have an information advantage over analysts with respect to FSI relative to SRI. Moreover, managers and analysts may bias their forecasts intentionally for the sake of personal gain, meaning that their forecasts may
395
not reflect their true information advantages. Hence, the question of whether or not managers have a greater information advantage over analysts with respect to FSI or SRI is an empirical issue which constitutes Brown and Zhou’s first research question. The results suggest that managers have a larger advantage over analysts in incorporating SRI into their forecasts rather than FSI. Their second research question compares the analyst forecasts made after management forecasts versus those made before, and examines whether the latter analyst forecasts show a greater improvement at incorporating past SRI rather than past FSI. They find that analysts’ forecasts improve more subsequent to managers’ forecasts by incorporating SRI versus FSI. Their results are consistent with analysts being aware of managers’ comparative advantage at integrating the information underlying returns into their earnings forecasts and using it to improve their post-management earnings forecasts accordingly. In a supplementary analysis, Brown and Zhou show that analyst errors are associated more closely with past SRI for firms with management forecasts than for those without, but they find no such differences for earnings changes or accruals, their two proxies for FSI. Their results are consistent with the view that managers are more likely to issue forecasts when analysts’ misinterpretations of SRI are greater, but not when analysts’ misinterpretations of FSI are greater. Their results strengthen the notion that managers possess a greater comparative advantage over analysts regarding SRI than regarding FSI, and are consistent with managers recognizing their comparative advantage and issuing forecasts when they are confident that analysts will misinterpret information. Overall, their evidence highlights the importance of SRI versus FSI in interactions between analysts’ and managers’ forecasts. Brown and Zhou make two major contributions to the literature. First, they show that managers have a greater information advantage relative to analysts in understanding how SRI impacts future earnings versus FSI. The prior literature has shown that analysts and managers are inefficient at incorporating public information into their earnings forecasts. Extant studies suggest that management forecasts are more accurate than analyst forecasts, but they have not identified the sources of managers’ information advantage. Brown and Zhou show that managers’ advantage over analysts is greater with respect to the information underlying stock returns than with respect to the information in financial statements. Second, their finding that analysts improve their forecasts more by incorporating SRI versus FSI after perusing management forecasts highlights how managers’ forecasts facilitate information dissemination to analysts. The literature (e.g. Bowen, Davis, & Matsumoto, 2002; Kimbrough, 2005) contends that managers’ voluntary disclosures help analysts to improve their forecasts, but it does not show the type of information that helps analysts. This paper reveals that managers help analysts more by improving their interpretation of SRI rather than FSI, thus furthering our understanding of the role of voluntary disclosures. Sheng and Thevenot, in ‘‘Quantifying Differential Interpretation of Public Information Using Financial Analysts’ Earnings Forecasts’’, examine analysts’ ability to capture
396
Editorial / International Journal of Forecasting 31 (2015) 392–398
the variation in opinion divergence using several analyses. Their paper contributes to the literature by separating the two possible explanations for investor disagreement following public disclosure: differences in prior beliefs and differences in the interpretation of public signals (cf. Lahiri & Sheng, 2008). More importantly, they use this decomposition to develop an improved measure of differential interpretation. The proposed metric is preferable to previously used proxies because of its strong alignment with the theoretical construct and its ease of implementation using any statistical software capable of regressions. The results provide convincing evidence that the proposed measure is as good as or superior to any previously used proxies. Specifically, the authors find that the new measure is related positively to trading volume, the informativeness of earnings announcements and the cost of capital, while it is related negatively to disclosure readability and management guidance precision. Furthermore, only the proposed measure provides statistically significant and consistent evidence in all empirical applications. In summary, their analyses indicate that the proposed measure captures the unobserved differential interpretation reliably in a variety of settings. Their study provides a better and widely applicable tool for a further understanding of the effect of public disclosure on investor behavior. In ‘‘Do Analysts Treat Winners and Losers Differently when Forecasting Earnings?’’ Jung, Pae, and Yoo use the I/B/E/S and CSRP data to examine whether the well-known positive association between past stock returns and analysts’ earnings forecast revisions differs for stocks that have experienced extreme positive or negative price changes. They show that: (i) a strong degree of asymmetry exists in the association between earnings forecast revisions and past stock returns, depending on whether a stock experienced a capital gain or loss; (ii) less experienced analysts tend to treat winners and losers differently, which leads to less accurate forecasts; and (iii) investors discount earnings forecast revisions made by analysts who treat losers differently, but the same does not seem to be true if the analysts treat winners differently. Furthermore, Jung, Pae and Yoo examine the effects of regulatory changes in 2002 on analyst forecast revisions, and find that analysts’ overall tendency to reflect past stock returns in earnings forecast revisions decreased after the Global Settlement, which is consistent with what recent studies have documented in the context of stock recommendations. However, more importantly, they report that analysts have continued to exhibit asymmetric patterns in earnings forecast revisions even after the Global Settlement, thus calling regulators’ attention to the asymmetry they document in the paper. Finally, the paper sheds light on the effects of the regulatory changes in 2002 on investors’ responses to analysts’ asymmetric behaviors towards winners and losers. Consistent with the stock market regulators’ ongoing efforts to educate investors, the market appears to start discounting the earnings forecast revisions issued by analysts who misinterpret the implications of winners’ and losers’ stock returns, because such behaviors result in poor forecast accuracies. In July 2012, McKinley Capital hosted a research seminar in Anchorage, AK, where researchers analyzed a
common Global Expected Returns (GLER) database that contained earnings forecasts and fundamental data, such as earnings, book value, sales, and debt variables for all globally-traded stocks with least two analysts over the period 1997–2011. The researchers were given the data and requested to produce portfolios that maximized the portfolio Geometric Mean. This database was shared with investment researchers, forecasters and university faculty members. An introductory analysis of the earnings forecasts in the GLER database is provided in the paper by Guerard, Markowitz, and Xu, ‘‘Earnings Forecasting in a Global Stock Selection Model and Efficient Portfolio Construction and Management’’, which extends the study by Bloch, Guerard, Markowitz, Todd, and Xu (1993) that documented the efficiency of fundamental-based stock selection models for Japanese and United States stocks. Stock selection models often use momentum, analysts’ expectations, and fundamental data. Guerard, Markowitz, and Xu find support for composite modeling using these sources of data, as well as evidence supporting the use of APT multi-factor models for portfolio construction and risk control, then incorporate earnings forecasts and price momentum into a stock selection model. Guerard, Markowitz and Xu add a composite earnings forecasting variable CTEF and a price momentum factor to the Bloch et al. model, to create a ten-factor stock selection model for global expected returns, which they refer to as the GLER model. Given concerns about both outlier distortion and multicollinearity, Bloch et al. (1993) tested the relative explanatory and predictive merits of various alternative regression estimation procedures: robust regression using the bi-square criterion to mitigate the impact of outliers, latent roots to address the issue of multicollinearity and weighted latent roots, denoted WLRR, and a combination of robust and latent roots. The predicted returns and the predictions of risk parameters were then used as inputs for a mean–variance optimizer to create mean–variance efficient portfolios in the global financial market. An increase in lambda serves to produce portfolios with higher geometric means (GM), Sharpe ratios (ShR), and information ratios (IRs). Guerard, Markowitz, and Xu report several sets of results generated by the traditional Markowitz mean–variance analysis. Portfolio risk is divided into systematic and unsystematic risk, or stock specific risk. Systematic risk is often measured as the the risk of the market, countries, sectors, and other factors, such as value, momentum, and size. Guerard, Markowitz, and Xu also test a practitioner’s variation that weights the APT-estimated systematic risk at three times the importance of the specific risk, with the optimum being when the weights are estimated to deviate from the market weights by two percent. The efficient frontiers of the three sets of portfolios report substantial excess returns for any given percentage of systematic risk once the level of risk approaches a lambda of 100 or portfolio tracking errors of at least 5%–6%. The authors estimate the FactSet-based GLER database and test the usefulness of the alpha alignment factor (AAF) in two applications. First, they report that GLER portfolios created using the Axioma world-wide statistically-based risk model and
Editorial / International Journal of Forecasting 31 (2015) 392–398
the Axioma world-wide fundamentally-based risk model, earned significant excess returns. This paper reports a large set of tracking error optimizations that reveals that the geometric means and Sharpe ratios increase with the targeted tracking errors; however, the information ratios are higher in the lower tracking error range of 3–6%, with at least 200 stocks in the optimal portfolios, on average. The authors find that statistically-based risk models using principal components, such as Sungard APT and Axioma, produce more efficient trade-off curves than fundamentally-based Axioma risk model using their variables. In the second AAF application, they construct portfolios using an earnings forecast variable, CTEF, as the expectedreturn model. They create CTEF-based portfolios using the Axioma world-wide statistical risk (STAT) and fundamental risk (FUND) models under conditions identical to those of the GLER model. Guerard, Markowitz, and Xu et al. report that portfolios constructed using the statistical (STAT) risk model dominate those constructed using the fundamental (FUND) risk model. The STAT model procedure increases the number of securities in the optimal portfolios substantially. The CTEF variables require more stocks than the GLER model in the Axioma portfolio simulations. The authors limit the number of securities to only 70 stocks each month and obtain a more investable solution that is still consistent with the risk-return tradeoff. As the tracking errors rise, the Sharpe ratios generally also rise. The information ratios tend to support the creation of portfolios using the lower tracking errors of the GLER model. Guerard et al. then develop and estimate three levels of testing for stock selection and portfolio construction. The use of multi-factor risk-controlled portfolio returns allows the authors to reject the data mining corrections test null hypothesis. The authors have since updated their research through to December 2013, and found similar results. Interested readers may request updated information from the corresponding author. The anomalies literature can be applied in real-world global portfolio construction. Shao, Rachev, and Mu, in ‘‘Applied Mean-ETL Optimization in Using Earnings Forecasts’’, examine the mean-ETL portfolio optimization based on the CTEF variable in the global market. The ARMA-GARCH model with the MNTS innovation is used to generate the out-of-sample scenarios for the portfolio optimization. The authors show that their approach with the CTEF variable can produce statistically significant active returns in the out-of-sample test, 1999–2011. Compared with previous works on mean-ETL portfolio with GLER and USER data, the active returns produced in the CTEF portfolio are driven more by stock selection. Xia, Min and Deng, in ‘‘Global Efficient Portfolio Construction Using Earnings Forecasts’’, analyze the performances of optimal global portfolios constructed based on the Guerard, Markowitz, and Xu CTEF forecast variable. Under the Markowitz mean–variance framework, applied optimization techniques are employed to obtain the optimal portfolios that satisfy practical requirements on risk-tolerance, the turnover rate, and the tracking error. Xia, Min and Deng’s empirical analysis demonstrates the effectiveness of the consensus temporary earnings forecast variable in selecting stocks for constructing
397
mean–variance optimal portfolios with significant active returns. In particular, the CTEF variable with a simple forecast based on historical data is capable of generating portfolios with active returns that are comparable to those generated by sophisticated multi-factor return forecast models. Gillam, Guerard, and Cahan, in ‘‘News Volume Information: Beyond Earnings Forecasting in a Global Stock Selection Model’’, report that earnings forecasts produce highly statistically significant asset selection, active equity, and total active returns over the period January 2004–December 2013. Information on news volume residuals enhances returns in the Boolean signal, information coefficient and top/bottom decile spread measures relative to the earnings forecasting variable, CTEF. Moreover, the information on news volume residuals enhances the geometric means, information ratios, and Sharpe ratios in global portfolios. There are portfolio returns beyond the earnings forecasting portfolio returns. Beheshti, in ‘‘A Note on the Integration of the Alpha Alignment Factor and Earnings Forecasting Models in Producing More Efficient Markowitz Frontiers’’, reports additional evidence on the risk and performance characteristics of several optimized portfolios that take advantage of the detailed knowledge of CTEF, an earnings forecasting expectations component of the GLER model of Guerard and Markowitz et al. (2013). Furthermore, he utilizes Axioma’s multi-factor statistical and fundamental global risk models for deriving covariance estimates, FactSet Research Systems’ analytical content sets for producing CTEF alpha signals, and Axioma’s Portfolio Optimizer through the FactSet Portfolio Simulation Engine for building the portfolios. Beheshti reports four conclusions: (1) the CTEF variable is highly statistically significant at identifying mispriced global securities; (2) portfolios constructed using the CTEF variable and an Axioma statistically-based risk model outperform portfolios constructed using the CTEF variable and an Axioma fundamentally-based risk model when the performance is measured by the geometric mean, the Sharpe ratio, and the information ratio; (3) the AAF-constructed portfolios dominate the non-AAF portfolios when performance is measured by the geometric mean, the Sharpe ratio, and the information ratio; and (4) targeting higher tracking errors produces higher geometric means and Sharpe ratios than targeting lower tracking errors. What have we learned from the use of forecasting models in the analysis of business cycles, foreign exchange, and corporate earnings? The empirical evidence from the studies in this special issue suggests that markets are not perfectly efficient. Excess returns can be earned in foreign exchange and corporate earnings. Business cycles can be predicted with appropriate errors at the margin. Simplistic models generally do not win. Simple earnings forecasting revisions do not generate excess returns, but more complex earnings forecasting models and the disagreement between analysts can be useful. Earnings forecasting drives portfolio outperformance, but it is not a product of simple models. Abnormal news volume can enhance returns. A common theme that runs through this collection of papers is the importance of forecasting in explaining complex global linkages in international finance, corporate earnings and macroeconomics.
398
Editorial / International Journal of Forecasting 31 (2015) 392–398
References Bowen, R. M., Davis, A. K., & Matsumoto, D. A. (2002). Do conference calls affect analysts’ forecasts?. The Accounting Review, 77(2), 285–315. Brown, L. D. (1993). Earnings forecasting research: its implications for capital market research. International Journal of Forecasting, 9, 295–320. Brown, L.D. (2008). Thomson Financial Research bibliography. Brown, L. D., & Guerard, J. B. (Eds.) (1998). International financial forecasting. International Journal of Forecasting, 14, 157–304. Special issue. Bloch, M., Guerard, J. B., Jr., Markowitz, H. M., Todd, P., & Xu, G.-L. (1993). A comparison of some aspects of the U.S. and Japanese equity markets. Japan and the World Economy, 5, 3–26. Guerard, J. B., Jr., Rachev, R. T., & Shao, B. (2013). Efficient global portfolios: big data and investment universes. IBM Journal of Research and Development, 57(5), Paper 11. Guerard, J. B., Jr., Markowitz, H., & Xu, G. (2013). Global stock selection modeling and efficient portfolio construction and management. Journal of Investing, 22(4), 121–128. Kimbrough, M. D. (2005). The effect of conference calls on analyst and market underreaction to earnings announcements. The Accounting Review, 80(1), 189–219. Levanon, G., Manini, J-C., Ozyildirim, A., Schaitkin, B., & Tanchua, J. (2015). Using financial indicators to predict turning points in the business cycle: the case of the leading economic index for the United States. International Journal of Forecasting, 31, 419–438. Lahiri, K., & Sheng, X. (2008). Evolution of forecast disagreement in a Bayesian learning model. Journal of Econometrics, 144, 325–340.
Ramnath, S., Rock, S., & Shane, P. (2008). The financial forecasting literature: taxonomy with suggestions for further research. International Journal of Forecasting, 24, 34–75. Teräsvirta, T. (2006). Forecasting economic variables with nonlinear models. In G. Elliott, C. W. J. Granger, & A. Timmermann (Eds.), Handbook of economic forecasting: Vol. 1 (pp. 413–457). Amsterdam: Elsevier. Ziemba, W. T., & Schwartz, S. L. (1991). Invest Japan. Chicago: Probus.
John Guerard ∗ McKinley Capital Management, LLC, Anchorage, AK 99503, United States E-mail address:
[email protected]. Kajal Lahiri University at Albany: SUNY, Albany, NY 12222, United States E-mail address:
[email protected]. ∗ Corresponding editor.