Do sophisticated investors use the information provided by the fair value of cash flow hedges?

In addition, Sect. 2.1 provides additional details on the use of and accounting for cash flow hedges. There we state the necessary conditions under which our hypotheses would hold and empirically test that those conditions hold for our sample firms.

Hirshleifer and Teoh (2003, p. 380) note that investors may incorrectly estimate the value of firms “if hedge profits are marked-to-market whereas the long-term business risk the firm is hedging is not marked-to-market.”

For examples of forward-looking disclosures that accompany firms’ management forecasts, see Sect. 4.1 and Appendix 3.

Fischer et al. (2009, pp. 538–542) provide a comprehensive illustration of the accounting for cash flow hedges.

We use the most recent information available in electronic form. Our research design requires 3 years of forward-looking analyst forecasts. Thus, although our sample ends in 2008, we are using information through 2011 in our empirical tests.

Financial firms are eliminated for two reasons. First, Bodnar et al. (1998), Makar et al. (2013), and Campbell (2014) do not include financial firms. Second, financial firms operate in a different regulatory environment compared to nonfinancial firms. Thus the relationship between the level of cash flow hedging gains and losses and analysts’ forecast errors may not be the same for financial and nonfinancial firms. However, all of our inferences are unchanged if we include financial firms in our sample.

This classification is stable over time. For 20 % of our sample, we investigate whether this classification changed in 2008 (the last year our tests examine). Out of 99 firms, the classification changes in only one (1.0 %), and this change was the result of a firm entering into an interest rate hedge for debt it acquired during the second year it appeared in our sample.

Appendix 1 of Campbell (2014) shows this mathematically as long as three assumptions are met. First, the price of the underlying item follows a random walk. We have shown this to be the case in Table 2. Second, the firm must hedge its underlying items on a rolling basis. For a random sample of 20 % of our sample, we find that 99 % hedge the same items in their first and last years in our sample. Since the average expiration of hedge contracts is <1 year, finding similar hedging at intervals of up to 8 years suggests that firms hedge their transactions on a rolling basis. Finally, it must be the case that the cross-sectional variation in the hedge ratio within an industry is relatively constant across short periods (i.e., from t − 1 to t). Although this assumption is not empirically testable due to incomplete firm disclosures, prior theoretical research suggests that firms’ hedge ratios depend on firm-specific and relatively time-invariant factors such as size, risk, and the delta and vega of the manager’s stock and option portfolio (Smith and Stulz 1985; Froot et al. 1993; Geczy et al. 1997).

Managers typically provide their year t + 1 earnings forecast when they announce earnings for year t (i.e., for calendar year-end firms, in January of t + 1) (Anilowski et al. 2007).

All mean values of the continuous variables presented in Table 3 are significantly different from zero at the 0.01 level.

All of these correlations are significant at the 0.01 level.

The variance inflation factors across all of our models were 2.3 or less, suggesting that multicollinearity is not a concern in any of our analyses (Kennedy 2003).

As noted in Sect. 2.2, unrealized cash flow hedge gains/losses capture firm-specific information about the effect of underlying price movements on future profitability. Because we are (at least partly) interested in across-firm variation, we do not include firm fixed effects in our models. Nevertheless, in untabulated results, we estimate all of our models throughout the paper with firm fixed effects in place of industry fixed effects, and all of our results continue to hold.

Our sample is comprised of 486 unique firms, and the data is not a balanced panel across all years in the sample. Given the small number of firms and years represented in our sample, it is not clear whether standard errors clustered by firm and year would provide statistically reliable results (Petersen 2009; Gow et al. 2010). Nevertheless, in untabulated results, we estimate all our models with standard errors clustered by firm and year, and all of our results continue to hold.

These coefficients imply that one dollar of unrealized hedging gains translates into reduced earnings of 10.3 and 8.7 cents in years t + 2 and t + 3, respectively. However, these coefficients are based on our multivariate models with a significant number of control variables. If we instead estimate a univariate model (as in Bradshaw et al. 2001), the coefficients on HEDGE for years t + 2 and t + 3 are −2.27 and −1.73, respectively. These coefficients are consistent with findings in Makar et al. (2013) and Campbell (2014) and suggest that firms significantly under-hedge their exposures (i.e., hedge well below 100 % of their future transactions), so that a one dollar gain on a cash flow hedge ultimately translates to lower future profits of more than one dollar.

If instead of actual earnings reported in I/B/E/S we use GAAP earnings reported in Compustat as our proxy for “actual” earnings, our results for Tables 4 and 5 are unaffected.

Since we do not know precisely when the current year unrealized cash flow hedge gains and losses affect future earnings, it is difficult to truly isolate an adjusted forecast error. For these tests, we make the assumption that 100 % of the gain/loss in the current year affects earnings in each of the following 3 years. Our focus is on the intuition of whether the adjusted forecast error is larger or smaller than the unadjusted forecast error.

For brevity, we reference the papers that motivate the selection of our control variables and refer readers to those papers for a full description as to how these control variables have been shown to affect analyst forecast errors. We selected these control variables after reviewing the totality of the prior literature on analyst forecasts. One might argue that we should also include the weighted average level of accruals based on recent evidence from Drake and Myers (2011). In untabulated results, we do this, and all of our results throughout the paper are unchanged.

Our inferences are unchanged if we use a fully interacted regression model rather than a partition regression approach.

In untabulated results, we replace the dependent variable of these regressions with management forecast errors (rather than analyst forecast errors). Consistent with management possessing information about the implications of unrealized cash flow hedge gains and losses, we find no association between management forecast errors and HEDGE in any period. However, unlike with analyst forecast errors, data limitations from First Call require that we calculate management forecast errors using a forecasted number from twelve months or less before the actual earnings release date. This restriction undoubtedly improves managers’ forecasting ability for t + 2 and t + 3 because they are making their forecasts at t + 1 and t + 2, respectively. Consequently, these results should be interpreted with caution.

For ease of exposition, we focus our discussion on the case where the firm has a calendar year-end. However, as noted in Table 9, our sample still includes non-calendar-year-end firms. In addition, we focus on analyst corrections of year t + 2 earnings forecasts in this section, but the results are similar if we use t + 3 instead.