Introduction - Fuqua School of Business



Global Asset Allocation:

Univariate Screening Factor: Change in Forecasted Earnings

Table of Contents

Table of Contents 2

TABLES 2

FIGURES 2

ALCHEMY INVESTMENTS TEAM 3

INTRODUCTION 3

DATA 3

HYPOTHESIS 3

METHODOLOGY 4

RESULTS 5

CONCLUSION 10

APPENDIX 11

Tables

Table 1: Summary of Performance Measures 6

TABLE 2: ENDING ANNUAL INDEX LEVEL 7

TABLE 3: NUMBER OF STOCKS PER CHANGE IN FY 1 EARNINGS 10

Figures

Figure 1: Ending Annual Index Level 8

FIGURE 2: OVERALL ENDING INDEX LEVEL 8

FIGURE 3: SPREAD RETURN OVER TIME 9

FIGURE 4: SPREAD RETURN VS. STANDARD DEVIATION 10

Alchemy Investments Team

Clark Cheng

Jeff Chung

Jungwoo Yang

Tara Ellwanger

Introduction

The final step in completing an asset allocation exercise, after the strategic and tactical asset allocation, is stock selection. There are many different stock selection models, two of which are multivariate regressions and sorting strategies. Both models use screening factors to isolate out those stocks that are most likely to outperform and underperform the benchmark index. The most important determinant of a successful stock selection strategy is deciding which screening factors to use

Past research has shown that factors representing future expectations are a valuable data source. In this research project, we attempt to measure the value of one such factor, change in FY 1 earnings forecasts. We use the I/B/E/S database of analyst forecasts to examine the monthly changes in this variable over a six month period. Our analysis yielded results confirming our belief that change in FY 1 earnings is a viable screening factor for a successful stock selection model.

Data

Our analysis was based on data obtained from two sources: I/B/E/S and Compustat/CRSP. I/B/E/S is a database beginning in 1976 of analysts’ forecasts on individual stocks. Compustat/CRSP is a detailed database of stock information, including monthly stock prices and holding period returns.

Hypothesis

Past research has shown that data containing investors’ expectations have tremendous value in selecting stocks. Further, research shows that expectations embedded in monthly changes to mean forecasted FY 1 earnings is a valuable predictive variable. Although past research has studied this factor, we believed that more analysis was possible to examine if more predictive power could be extracted by looking a trend to changes in mean forecasted FY 1 earning. Hence, our analysis looks at how many positive changes have occurred relative to negative changes over the prior six month period. We expect that those stocks with a greater number of positive changes relative to negative changes will yield higher returns the following month. In our analysis, we do not examine the degree of change to forecasted earnings nor do we examine how recent the change occurred over the past six months. We hypothesize that both factors could yield valuable information; however, in ignoring this extraneous data, we believe we are filtering out noise in the dataset. We believe that our analysis represents a base case to understanding how trends in changes to forecasted earnings are a valuable data source.

Methodology

The amount of data and work necessary to explore the value of this screening factor was tremendous. The I/B/E/S database alone was over 80 Megs of text and contained 700,000 rows of data. In comparison, the Compustat database was over 120 Megs of text and contained over 1.1 million rows of data. The most difficult aspect of analyzing this data was finding enough computing power to run the models. The difficulty arises because of the way the data is organized. Every forecast for every month uses one row of data with many repeating variables. For example, company name, ticker, CUSIP, Permno, SIC and etc. will be repeated for every monthly forecasted earnings entry.

Excel is the benchmark program for running financial models in business school; however, both datasets go well beyond Excel’s capabilities. As a result, we needed to come up with a more innovative and resourceful solution. It should be noted that possibly a more powerful statistical software program, such as SAS or SPSS could run the dataset fairly easily, but given our resources, we were limited to Microsoft Office.

Our first step was to use the relational database power of Access. After importing both datasets into Access, we used the table analyzer to break apart the data into multiple tables. Without breaking apart the datasets, it takes hours for Access to run simple filters or sorts. After breaking apart the data, relationships were created to link both datasets. The key link between both datasets was Permno. Permno is a unique identifier used by different data compilers to identify stocks. Other identifiers include CUSIP and Ticker; however, not every data organization uses these identifiers consistently. Our dataset reconciles the Permnos between both data compliers to ensure accuracy and reliability.

After breaking apart the datasets and creating relationships, queries were formed to sort out the relevant dataset for our analysis. The first query recompiled the I/B/E/S dataset with certain filters. The main parameter we set was the period of time we wanted to run the analysis over, which was 1990 to 1999. The next query we ran was a cross tab query. This query is designed with Permnos running down the left side column, months running across the top row and the associated forecasted mean FY 1 earnings in intersections. We ran another identical query with monthly holding period returns as the associated intersection. Both queries were then exported into an Excel format.

After importing the queries, Excel was able to run the remainder of the analysis. However, the Excel workbook containing both datasets is still extremely large, but at least within Excel’s limitations. First, with the forecasted earnings worksheet, we calculated the percentage change between months. Next, we wrote a formula calculating the number of positive changes and subtracting the number of negative changes over a six month period. At this point Excel will run more efficiently if you copy and paste the worksheet as values rather than formulas. So, at this point we have a worksheet with Permnos running down the left side column, month running across the top row, and the number of positive less negative, monthly mean FY 1 earnings revisions in the intersections.

Sort the worksheet in descending order by change in FY 1 earnings for every month and copy and paste the left side list of Permnos to another worksheet. A simple VBA program will do this fairly quickly. As a check, you should have a list of Permnos in descending order according to the difference in changes to FY 1 earnings. The top Permnos listed should have a maximum of six positive changes and the bottom Permnos, a maximum of six negative changes.

The final table to create is a simple VLOOKUP table. The VLOOKUP formula will take each Permnos and find its associated monthly holding period return. It is important to move forward the holding period returns to the following month. Remember, for every month we have the past changes in FY 1 earnings, including the current month. Once we know this information, we can create portfolios and calculate returns for the following month. We cannot estimate returns for the month that we are still calculating changes in FY 1 earnings because this information may not be ready until the end of the month. Running this formula for every Permnos listed in order of their change in FY 1 earnings will yield a table from which to create portfolios. This final table will list the monthly holding period returns in descending order by changes in FY 1 earnings.

Finally, we can create portfolios to examine whether changes in the trend of FY 1 earnings is a valuable screening factor for stock selection. We examined portfolios containing 100 stocks and 50 stocks. Both yielded similar results, with the 100 stocks outperforming slightly. Remember, the higher the number of stocks, the higher the turnover and associated trading costs. Additionally, the higher the number of stocks, the more diversified the portfolio. Conversely, the higher the number of stocks, the greater the likelihood that the portfolios will contain stocks with a smaller number of positive changes relative to negative changes in FY 1 earnings. These are all issues that need to be carefully considered.

After creating portfolios of the top 100 and bottom 100, we measured the outperformance and underperformance of each portfolio on a number of measures.

Results

The following table summarizes our results.

Table 1: Summary of Performance Measures

[pic]

In the final analysis, we looked at 109 monthly portfolio returns between August 1990 and September 1999. The results presented are based on monthly equal weighted portfolios. Conversely, we could have examined annual value weighted portfolios. Further, we only publish results from the top 100 and bottom 100. It would also be informative to look at other fractiles between our portfolios. Finally, we assumed that we would create a spread portfolio by buying the top 100 and shorting the bottom 100. All results are compared to the S&P 500 as our benchmark.

From a monthly return per risk perspective, the spread portfolio has a higher return given a lower monthly standard deviation. Although other portfolios may give higher returns per risk, including the S&P 500, the spread portfolio should have an overall lower risk profile. For example, even though, the monthly standard deviation may be the same for a given amount of return, the portfolio could be reducing other risk factors such as systematic market risk. Comparing the benchmark S&P 500 portfolio with a beta of 1, to the spread portfolio with a beta of .1, suggests that other risk factors need to be considered when making investment decisions. Conversely, the spread portfolio may have other risk factors relative to the S&P 500, such as interest rate risk or industry specific risk.

In evaluating the performance between the long portfolio relative to the short portfolio, the long portfolio outperforms as expected. The monthly absolute outperformance between the two portfolios is 8 basis points. On an excess returns basis, the delta obviously remains the same, but the relative starting point changes. Hence, the short portfolio has negative excess returns relative to the long portfolio by the same 8 basis points. Comparing the long portfolio to the S&P 500, suggests that the screening factor may not be as valuable a screening factor with the long portfolio as the short portfolio. It could simply be a better predictor of poor stock performance rather than good stock performance. Hence, we could simply buy the S&P 500 futures while shorting our bottom 100 portfolio. However, we must consider other risk factors such as beta in forming these spread portfolios. For example, what we just proposed may yield higher returns per standard deviation, but it may also take on additional risk factors such as beta. Overall, each of our portfolios yield a positive alpha, with the spread portfolio yielding a monthly alpha of 2 basis points. Given a beta of .1, an alpha of 2 basis points may be reasonable.

In terms of correct direction count, the long portfolio was accurate 62% of the time whereas the short portfolio was correct 44% of the time. This result is expected because in the short portfolio we are essentially betting that it will underperform, hence, it should have a lower correct direction count. Our portfolios were more accurate on the direction count when the S&P 500 was positive than negative. Overall, the spread portfolio showed promising results on the accuracy of the direction count.

Finally, in evaluating the relative returns on an annual basis, the results are extremely promising. The following table shows the relative annual index level at the end of each year starting with 100 in January. In every case, the long portfolio outperforms the short portfolio as expected. Overall, the spread portfolio is always positive, a good sign that our screening factor is value added.

Table 2: Ending Annual Index Level

[pic]

This chart shows the same information graphically.

Figure 1: Ending Annual Index Level

[pic]

The following chart shows the final index level assuming we invested $100 on August 1990. Based on the chart, it is clear that the long portfolio outperforms the short portfolio, with the spread portfolio performing in the middle with less risk exposure. Although the S&P 500 has a higher return, remember that it also has a higher associated beta.

Figure 2: Ending Time Horizon Index Level

[pic]

This graphs shows the monthly returns from the spread portfolio. It is positive 65% of the 109 months.

Figure 3: Spread Return Over Time

[pic]

This graph shows the positive relationship between spread return and monthly standard deviation. Beta risk for this portfolio is .1.

Figure 4: Spread Return vs. Standard Deviation

[pic]

This table shows the number of positive and negative FY 1 earnings revisions per year. Essentially, we are buying the stocks on the left side and shorting the stocks on the right side. Although we only looked at the top 100 relative to the bottom 100, notice that the number of stocks within each column differs. Hence, if the top 100 ended in the 5 positive changes in FY 1 earnings column, it is completely random which stocks were included in our portfolio out of all the stocks with 5 positive changes in FY 1 earnings. Further research should consider a more insightful strategy for selecting stocks within each cutoff.

Table 3: Number of Stocks Per Change in FY 1 Earnings

[pic]

Conclusion

Overall, our research show promising results. As expected, the use of expectations data is useful in screening out stocks. From a naïve perspective, the simple one month change in FY 1 earnings may yield positive results; however, a more in depth carefully conceptualized approach will yield even more powerful consistent results. Consequently, we believe that we have a valuable input to the stock selection phase of a global asset allocation exercise.

The next step would be to consider four other possible screening factors. In scoring each factor in a typical sorting strategy, we would simply enter the rolling five year average mean returns and standard deviations into a Black-Litterman optimization program to find monthly dynamic weights for each factor. Finally, after sorting each stock by its total score, we would create our spread portfolio. This will be our index portfolio which would represent the U.S. equities market. By using this proprietary index as our U.S. equity asset class, we can reoptimize the strategic and tactic asset allocation, bringing us back full circle and completing an internally consistent asset allocation exercise.

Appendix

More details of our analysis can be downloaded off this website. Note, the Access file is over 865 Megs and the Excel file is over 80 Megs, however, a more simplified Excel workbook is provided on this website. Further questions regarding this analysis should be forwarded to Clark Cheng.

................
................

In order to avoid copyright disputes, this page is only a partial summary.

Google Online Preview   Download