How Will the Court Decide? – Tax Experts’ versus Laymen's Predictions

Abstract

We conduct a survey of German tax professionals (tax advisors and revenue agents) and laymen to examine whether tax experts more accurately forecast the outcomes of five real cases from the German Federal Fiscal Court. With an average of 2.39 correct predictions among experts and an average of 2.49 correct predictions among laymen, our results reveal no significant difference in forecasting accuracy between the two groups. Additionally, neither general nor task-specific tax expertise increases the experts’ forecasting accuracy. This unpredictability of tax court decisions indicates that accounting rules and taxpayer penalties that rely on accurate predictions of tax court decisions may need to be re-evaluated. Moreover, our results indicate the existence of two types of ‘advisor bias’. First, tax advisors exhibit a significantly higher level of overconfidence in comparison to other experts (i.e. revenue agents) and laymen. In particular, they believe that they correctly predict, on average, 1.52 more cases than they actually do. Second, we find some evidence indicating that tax advisors acting as client advocates form stronger appeal recommendations than revenue agents.

JEL Classification:

• M40
• K20
• H20

Acknowledgments

We are grateful to Laurence van Lent (editor), two anonymous reviewers, Martin Fochmann, and Axel Möhlmann for their helpful comments and suggestions. We also thank Benjamin Peuthert and Louise Schnitzer for supporting the data collection.

Supplemental Data and Research Materials

Supplemental data for this article can be accessed on the Taylor & Francis website, doi:10.1080/09638180.2015.1114423

Notes

1 The characteristics variables are circuit of origin; issue area of the case; petitioner type (e.g. the United States, an employer); respondent type; ideological direction (liberal or conservative) of the lower court ruling; and whether the petitioner argues that a law or practice is unconstitutional (Ruger et al., 2004, p. 1154).

2 Except for the second case, the articles do not include personal judgments regarding the case or individual predictions about the outcome at the Federal Fiscal Court that may affect subject predictions. In the second case, the author explicitly states that he agrees with the lower court decision. This could influence tax professionals forecasting decisions. Note however, that this would bias tax professionals’ opinion in the correct direction.

3 In the questionnaire for revenue agents, we ask the agents for the strength of a recommendation against a pro-taxpayer decision.

4 We additionally conduct all following analyses based on the complete data set without excluding these cases. Main results remain qualitatively unchanged.

5 The survey also allows respondents to select ‘other’ reasons as the basis for decision (see Table 2). Examples include intuition, gut instinct, sense of justice, interpretation of the provided extracts of tax law, logical reasoning, and legal knowledge. The answers can be broadly classified into three categories: 1, intuition (about 20 % of all answers); 2, legal knowledge (about 55 %); and 3, logical reasoning (about 5 %).

6 With respect to the outcome of the proceedings, two of the five proceedings (cases 3 and 5) are judged in the taxpayer's favor. According to the annual reports of the German Federal Fiscal Court for 2011, 2012, and 2013, about 40% of all revisions were in the taxpayer's favor. This matches the outcome of our sample exactly. Thus, in this regard, our sample seems to be quite representative. However, all five decisions of the local tax courts are confirmed by the Federal Fiscal Court, whereas on average only 50 % of the local tax court decisions in 2011–2013 were confirmed. With respect to this issue, our sample seems to be rather untypical. However, we do not see how this biases our results, as knowledge about a 50–50 chance should neither improve nor deteriorate forecasting accuracy.

7 In case 4, three different outcomes are possible, thus the expectation is given by

8 Alternatively, we aggregate these categories to two groups (having more or less than 10 years of professional work experience) and perform a t-test to compare means. Again, no significant difference is found (p = .555).

9 We additionally perform non-parametric tests for all mean comparisons provided in Table 4. Results (not reported) remain qualitatively unchanged.

10 For each regression specification and all following multivariate analyses, we test independent variables for multicollinearity by means of variance inflation factors and pairwise Pearson's correlation coefficients obtained from the pooled data. No multicollinearity problem is evident. In this and all following regression models, we prefer (ordered) logit specifications over (ordered) probit specifications, as they offer an easier interpretation of effect sizes as log odds/odds ratios or a lower Akaike's information criterion. Main results obtained from (ordered) probit specifications remain qualitatively unchanged.

11 Alternatively, we include an indicator variable for each category of professional work experience (3–5 years, 5–10 years, 10–20 years, more than 20 years). Results remain qualitatively unchanged. However, we refrain from presenting those results, because some experience variables are only very rarely occupied and variance inflation factors from the pooled data raise collinearity concerns.

12 To test that revenue agents and tax advisors differ in forecasting accuracy, we perform a Wald test on the equality of their coefficients in the fitted model. Results indicate that coefficients significantly differ (Wald test, p = .0303).

13 Note, however, that this difference becomes insignificant in one of our robustness test (the subsample analysis excluding cases predicted by tax professionals with no relevant specialization) and thus should be interpreted with caution.

14 As previously mentioned, the experience variable equals 1 if the tax professional has more than ten years of professional work experience. In additional tests, we use alternative experience dummy variables based on thresholds of 3, 5, and 20 years. Results (not reported) show that neither variable significantly increases forecasting accuracy. With respect to the 20-year threshold, we find a negative correlation in some model specifications. Experience, as measured by years of professional work, does not increase forecasting accuracy.

15 Odds are defined as the ratio of the probability of belonging to one category of the dependent variable and its converse probability (i.e. the probability of belonging to any other category of the dependent variable).

16 First, we simultaneously include dummy variables for each experience category (3–5 years, 5–10 years, 10–20 years, more than 20 years). The results (not reported) indicate that no experience level significantly affects overconfidence. Second, we separately include dummy variables based on 3-year, 5-year, and 20-year thresholds. Results (not reported) suggest that tax professionals with more than 20 (less than five) years of work experience show a significantly higher (lower) level of overconfidence. Thus, tax professionals might overestimate the effect of experience on their own performance.

17 The proportion test does not account for within-subject correlation. We thus complement the univariate test by using a binary regression, with standard errors clustered at the individual level, to capture intra-subject correlation. Results (not reported) remain qualitatively unchanged.

18 Again, we additionally perform a non-parametric mean comparison test. Results (not reported) remain qualitatively unchanged.

19 Detailed regression results are available upon request.