Evaluating the information content of earnings forecasts

Abstract

This study develops a framework to compare the ability of alternative earnings forecast approaches to capture the market expectation of future earnings. Given prior evidence of analysts’ systematic optimistic bias, we decompose earnings surprises into analysts’ earnings surprises and adjustments based on alternative forecasting models. An equal market response to these two components indicates that the associated earnings forecast is a sufficient estimate of the market expectation of future earnings. To apply our framework, we examine four recent regression-based earnings forecasting models, alongside a simple earnings-based random walk model and analysts’ forecasts. Using the earnings forecasts of the model that satisfies our sufficiency condition, we identify a set of stocks for which the market is unduly pessimistic about future earnings. The investment strategy of buying and holding these stocks generates statistically significant abnormal returns. We offer an explanation as to why this and similar strategies might be successful.

Keywords:

• earnings forecasts
• earnings response coefficient
• market expectation of future earnings
• portfolio selection
• analysts’ forecasts

JEL codes:

• G11
• G17
• M40
• M41

Acknowledgements

We thank Edward Lee, Mark Clatworthy, Nikola Petrovic, Sheila Ellwood, Colin Clubb, Daniela Acker, Christian Leuz, Beatriz García Osma, William Rees, two anonymous reviewers and participants at the 2015 European Accounting Association Doctorial Colloquium for helpful comments and suggestions. All remaining errors are our own.

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

1. Ohlson (1991) comments: ‘Another theoretical problem concerns the relevance of unexpected earnings as a variable explaining returns. This construct appears to have the status of a “folklore concept” with limited economic content’.

2. The former error equals the deviation of analysts’ forecast from the actual reported earnings while the latter is the deviation of model-based forecasts from analysts’ forecast.

3. The HDZ and SO models base their earnings forecasts purely on accounting information. The KLSS model uses both market information (past stock returns) and analysts’ forecasts. The HW model is adapted by Harris and Wang (2013) using the theoretical approach in Ashton and Wang (2013). This model uses both accounting and market information to predict future earnings.

4. Because all empirical work concerns the aggregate behaviour of the variable across firms, at this stage firm subscripts are omitted.

5. is not strictly observable yet could be implicitly studied through the stock price reaction during earnings announcements.

6. We adopt the usual mathematical convention that capitals denote random variables and the corresponding lower-case letters the realization of those variables.

7. For reasons of clarity, at this stage time subscripts are omitted.

8. The unitalicized, unsubscripted followed by {} in equations (3) and (4) denotes the expectations operator.

9. Hou et al. (2012) generate forecasts at end of June for all firms with different fiscal year-ends and compare these with the latest analysts’ forecast. Hence, for firms with a fiscal year-end of July, their forecasts and analysts’ forecast are just a month away from the actual earnings announcement, yet they are then matched with annualized abnormal returns for model valuations. Their results are subject to this mismatch between the return window and the horizon of the earnings surprise measure. In our study, forecasts of one-year-ahead earnings are generated in April, so are approximately a year from the next annual earnings announcements. Earnings surprises associated with these forecasts are then matched to annualized abnormal returns at the model-valuation stage.

10. Hou et al. (2012) (who use all firms with a sample period from 1968 to 2008) have a mean income before extraordinary items (IB) of $49.07m and mean total assets (AT) of $1529.78m. Konchitchki et al. (2013), who use December-fiscal-year-end firms with a similar sample period (1985–2010), have a mean change in EPS deflated by stock price (CIB) of 0.003, a mean lagged one-year return (RET) of 0.194, and a mean analysts’ forecast of changes in EPS deflated by stock price (CAF) of 0.019. Harris and Wang (2013) (1963–2011) report a mean earnings per share (IBPS) of $0.685, a mean book value per share (B) of $9.105, and a mean adjusted stock price (APRC) of $13.51. So (2013) does not present summary statistics.

11. The earnings surprises in the HDZ model are scaled by market capitalization, while those in the AF, RW, HW and SO models are scaled by adjusted stock prices to produce earnings surprises in the form of forward earnings yields. Since the KLSS model predicts changes in earnings per share deflated by price, there is no need to further scale earnings at this stage.

12. The AESs in analysts’ forecast reported are further disadvantaged by the ‘optimism’ of 0.0146 in Table 4. The triangle inequality implies that the absolute deviation is between the reported figure 0.0297 and 0.0151 (=0.0297–0.0146) in the absence of such bias. In addition, the high value of the t-statistics suggests less variation in the absolute deviation than in the other measures.

13. A negative entry in a column implies that the column variable has a lower absolute error, while a positive entry in a row implies that the row variable has a lower absolute error.

14. Different models predict different measures of earnings, which prevents an easy comparison of ERCs. Hence, to enable a direct comparison, we follow Brown et al. (1987a) and Hou et al. (2012) by standardizing all proxies for earnings surprises each year to achieve unit variance. This has the added advantage in our case of making the resulting coefficients directly interpretable in terms of the theory developed in Section 2.

15. We adopt the approach used by Hou et al. (2012) to estimate the unexpected returns (UR) in equations (4) and (5), which we refer to as in equation (12).

16. On checking, the correlation coefficients in Tables 7 and 8 and the observed values of the weights are found to be consistent with their theoretical relationships established in Appendix A (equation (A7)).

17. Our approach is distinct from that of So (2013) since we base our investment strategy on a proxy of expected market adjustment for analysts’ forecast errors and mispredictions.

18. When we apply the Fama and French (1993) three-factor model, the corresponding results are 3.7%, 7.9%, 10.9% and 12.36%.

19. This is to make the earnings figure from COMPUSTAT compatible with the I/B/E/S earnings per share (Bradshaw and Sloan 2002). The tax rate is assumed to be 35%.

20. The accruals equal the change in current assets plus the change in debt in current liabilities minus the change in cash and short-term investments and minus the change in current liabilities.