Climate Disasters and Analysts’ Earnings Forecasts: Evidence from the United States

Abstract

We examine the relationship between climate disasters and analysts’ earnings forecasts in the United States. We find that climate disasters are associated with deteriorated analyst forecast properties proxied by forecast errors and forecast dispersion. We reason that the volatility of return on assets and of cash flows, and lower financial statement comparability, are three potential channels through which climate disasters influence analyst forecast properties. We also find that this relationship is more pronounced for firms in climate-vulnerable industries. Results from the market reaction tests further support our main findings by showing that the stock market responds less strongly to positive earnings surprises during periods of high climate disasters. Our results are robust to a battery of sensitivity tests, including a two-stage least squares approach and a difference-in-differences specification. Overall, the results shed light on the association between climate disasters and analysts’ earnings forecasts, which has significant implications for academics, investors, and standard setters.

Keywords:

• Climate disasters
• Climate change
• Analyst forecast errors
• Analyst forecast dispersion

JEL codes:

• G40
• Q54
• M21

Acknowledgment

We thank Maria Correia (Editor) and two referees for their invaluable insights and suggestions. Kanagaretnam also thanks Social Sciences and Humanities Research Council (SSHRC) for its financial support.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Supplemental Data and Research Materials

Supplemental data and research material are available in an online Supplement at the journal’s Taylor and Francis website, https://doi.org/10.1080/09638180.2024.2364785.

Appendix OA1. Additional descriptive statistics

Appendix OA2. Endogeneity tests

Appendix OA3. Robustness tests

Appendix OA4. Definitions of variables used in endogeneity and robustness tests

Figure OA1. Trends in climate disaster damages in the U.S. for 2001-2017

Table OA1. Sample distribution by state and year

Table OA2. Cumulative climate disaster property damages by state for 2001-2017

Table OA3. Endogeneity tests

Table OA4. A PSM-DiD analysis

Table OA5. Robustness tests

Notes

1 According to prospect theory, investors have a greater aversion toward losses than equivalent gains, indicating a stronger response to losses than to gains, which is supported by our asymmetrical findings.

2 Oh and Oetzel (2022) argue that climate change uncertainty is highly unpredictable and unknowable because of a mixture of natural forces and human involvement. In contrast, natural forces play a much less important role in other uncertainties, such as economic policy or political uncertainty.

3 Prior literature suggests that analysts may be unlikely to be distracted under the influence of natural disasters. Murphy and Smith (2015) find that analysts will likely escape being inattentive since they are trained to deal with chaos and multitask. They also find that analysts immerse themselves in their work and have little free time. Groysberg and Healy (2013) demonstrate that analysts are typically supported by their brokerages that outsource certain works to increase their attention capacity. Last, analysts normally cover related firms. That is, they can apply the information garnered from one firm to another, which increases the efficiency of their work.

4 There is no universally accepted definition of uncertainty. Therefore, we follow prior literature (e.g., Bloom, 2014) and define it as the difficulty in forecasting the likelihood of unknown outcomes.

5 There is also a vast stream of literature documenting that analysts are, on average, optimistically biased to achieve certain goals such as increasing upward mobility and pleasing the management (e.g., Abarbanell, 1991; Lim, 2001). Although analysts are less likely to be penalized for issuing optimistic forecasts during high uncertainty periods, continuous opportunistic behaviors are constrained by career and reputational concerns.

6 Given that our sample is collected in the post-Reg FD period, analysts have less incentive to issue optimistic forecasts to please the firms they cover to obtain private information. However, we acknowledge that there is still some room for them to obtain some private information from managers, as suggested by prior studies.

7 Prior research finds that Institutional Investor-ranked analysts possess higher ability as evidenced by higher forecasting accuracy, stock recommendation profitability, and report readability (e.g., De Franco et al., 2015; Stickel, 1990, 1992).

8 Reg FD was passed in 2000 to prohibit firms from disclosing private information to market participants such as security analysts. Given that the private information disclosed by managers to analysts may influence their forecast accuracy as indicated in the prior literature, limiting our sample to the post-Reg FD period can eliminate the distortion caused by private disclosures.

9 We obtain a firm’s headquarters location using the data and code provided by Gao et al. (2021) since Compustat does not report the historical locations of firms’ headquarters.

10 As pointed out by Gall et al. (2009), one potential limitation of the SHELDUS data is that economic losses are equally distributed across counties if they are simultaneously affected by a climate event. Put it differently, using the county-level data involves the tenuous assumption that all counties are equally affected by the same climate disaster. However, adopting a state-level climate disaster measure can avoid this potential drawback. As a robustness test, we also replicate our regression using the county-level measure and the results (which are reported in Panel A of Table OA5 in the Online Appendix) are qualitatively unchanged.

11 As mentioned earlier, a reason that we focus on climate disasters rather than geophysical disasters is that prior studies suggest that the impact on economic activities of these two types of disasters may differ (e.g., Klomp, 2017; Skidmore & Toya, 2002). Thus, our study covers a total of 13 types of climate-related events, including coastal events, drought, flooding, hail, heat, hurricane, landslide, lightning, severe storm, tornado, wildfire, wind, and winter weather. We exclude geophysical natural hazards, such as earthquakes and volcanic eruptions, from the analysis, and in fact, no such type of major disaster occurred in 2001–2017 in the mainland U.S.

12 Although employing a single index can potentially camouflage the heterogeneity in the effects of disparate disasters, arguably, the total economic damages is one of the most important indicators in evaluating the severity of natural disasters. For example, in Eckstein et al.’s (2019) global climate risk index (including the U.S.), economic losses and the number of deaths are the main factors in constructing the index. In addition, unlike Miao et al. (2018) who focus on the total of crop and property damages when investigating the dynamic fiscal response to natural disasters, we focus only on property damages in our baseline regressions because firms are mostly concerned with economic damages to property rather than crop damages. However, as a robustness test, we also construct a new measure that includes both property and crop damages.

13 Table OA2 in the Online Appendix shows the climate disasters proxied by the total annual climate disaster property damages (in millions of US$) at the state level during 2001-2017. We also provide a snapshot of the annual total property damages caused by weather-related disasters from 2001 to 2017 in Figure OA1 in the Online Appendix.

14 For detailed definitions of each variable we use in this study, see the Appendix.

15 Untabulated results suggest that our findings continue to hold when we control for uncertainty arising from infectious disease using the Infectious Disease Equity Market Volatility Tracker data from Baker et al. (2020).

16 Our illustration of economic magnitude is based on the results we report in Column (3). The impact of a one standard deviation increase in climate disaster risk on analyst forecast errors is calculated as 1.0e-11*0.830 (coefficient reported in Table 2 Column (3)) * 2811.74e+06 (standard deviation of climate disaster risk as reported in Table 1)/0.765 (mean of forecast errors as reported in Table 1) = 3.05%.

17 We drop year-fixed effects because they absorb the effect of these uncertainty measures (i.e., uncertainty is year-specific).

18 Our illustration of economic magnitude is based on the results in Column (9). The impact of a one standard deviation increase in climate disaster risk on analyst forecast dispersion is calculated as 1.0e-11$\ast$0.013 (coefficient reported in Table 2 Column (9)) * 2811.74e+06 (standard deviation of climate disaster risk as reported in Table 1)/0.008(mean of forecast dispersion as reported in Table 1) = 4.57%.

19 Our results are qualitatively unchanged when we include industry fixed effects in the model.

20 Prior literature also suggests that firms may engage in upward earnings management when facing climate change uncertainty, see, for example, Ding et al. (2021).

21 According to the literature, for the years prior to 2008, it is plausible to hand-collect these data from the Nelson’s Directory of Investment Research (NDIR). Post-2008 location is difficult to collect because NDIR has stopped updating since 2008. Therefore, we take an indirect approach to circumvent this obstacle.

22 The data are obtained from https://www.zippia.com/advice/best-cities-for-finance-analysts/. These cities are New York, NY, Boston, MA, Chicago, IL, Los Angeles, CA, Dallas, TX (Top 5), Houston, TX, San Francisco, CA, Atlanta, GA, Charlotte, NC, and San Jose, CA.

23 The estimated coefficients on the control variables are not reported for brevity.

24 Analyst consensus mean earnings forecasts are constructed based on the average earnings forecast issued by each following analyst. Untabulated results show that our findings are robust to (1) using a number of alternative windows such as (-3, 3), (0, 1), and (0, 3); and (2) using the median earnings forecast to calculate the earnings surprises.