Do stock markets react to soccer games? A meta-regression analysis

ABSTRACT

This study applies meta-regression analysis to aggregate a sample of 1126 empirical estimates of the stock market reaction to soccer matches collected from 37 primary studies. Our results indicate that winning a match is not associated with significant return effects for both national teams and individual clubs. In the case of lost matches, we find strong evidence for publication bias, i.e. negative returns are systematically overrepresented causing a biased picture of the true soccer match effect. After correcting for this bias, the mean return after losses by national teams becomes statistically insignificant and accounts for only $-5$ basis points. In the case of individual clubs, the corrected impact of a loss is a significant $-39$ basis points effect. In a further analysis, we identify various aspects of study design like regional differences, time period under examination and the design of empirical analysis to be responsible for the wide variation in previous study outcomes. Overall, our findings provide evidence against the hypothesis that stock markets are driven by sports sentiment in the case of national teams. Due to the existence of strong asymmetry in the returns after wins and losses of individual clubs, behavioural explanations cannot be fully ruled out.

KEYWORDS:

• sports sentiment
• soccer
• stock returns
• meta-analysis
• publication bias

JEL CLASSIFICATION:

• C83
• G14
• L83

Acknowledgements

We would like to thank Mark Taylor (the Editor), two anonymous referees, Ivo Welch, and Stefan Stöckl for their valuable feedback and suggestions. We also express our gratitude to participants at the British Accounting & Finance Association Annual Conference in Edinburgh, April 2017, and seminar participants at the Institute for Materials Resource Management, University of Augsburg. All remaining errors are our own.

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

¹ Ashton, Gerrad and Hudson (2003) find a strong association between the performance of the English national soccer team and daily changes in the FTSE 100 index.

² It should be noted that we use the term ‘publication bias’ to refer to selective reporting independent of a study’s publication status. As unpublished work like manuscripts and conference papers are written with the aim to get published, there is no reason to assume that the risk of selective reporting is lower in unpublished studies.

³ For example, lagged returns to control for first-order serial correlation, controls for day-of-the-week effects, or controls for non-weekend holidays.

⁴ In contrast to wins and losses, it is unclear how draws impact rational decisions or investor sentiment. Thus, we follow previous primary studies and do not analyse the return reactions after drawn matches (among others, Edmans, Garcia, and Norli, 2007; Kang and Park, 2015).

⁵ For the literature search and the subsequent meta-analysis, we follow best practices for MRA research issued by the Meta-Analysis for Economics Research network (Stanley et al., 2013).

⁶ Our search command consists of a combination of keywords related to soccer (soccer, football, sporting result, sport sentiment) and stock returns (stock return, stock price, economic impact).

⁷ The complete list of excluded studies is available on request from the authors.

⁸ Picking only one or a few estimates from each study (e.g. the ‘best-set’ or the ‘average-set’ of estimates) causes additional biases, requires objective selection rules to decide which estimate to prefer and leads to a loss of information about within-study variation (Stanley and Doucouliagos, 2012).

⁹ Due to confounding events, lagged return effects are difficult to measure for the impact of national teams’ results on broad stock indices. Therefore, lagged effects are only available in studies examining individual soccer clubs.

¹⁰ We use funnel plots only for a first visual indication about selective reporting. Inferences about publication bias are drawn from statistical testing of publication bias.

¹¹ We do not apply the fixed effects on Equation (3), because some of the moderator variables $Z_{ijk}$ are constant within studies and thus would be perfectly correlated with individual study dummies in the fixed effects model.

¹² It should be noted that also the expected sign of the moderator variables might be different for the two subgroups.

¹³ The Fédération Internationale de Football Association is the world soccer association.

¹⁴ None of the studies in our sample include matches of the other FIFA continental cups (Africa Cup, North America Cup or Nations Cup).

¹⁵ The selection of the top soccer nations follows Edmans, Garcia and Norli (2007), who analyse seven nations as top soccer countries. Due to missing observations for individual clubs from Argentina, Brazil or France, we had to exclude three of the seven nations from the classification.

¹⁶ Knock-out games are defined as games after the group session in national cups like the FIFA World Cup.

¹⁷ Primary studies typically measure market expectations by the implicit probabilities observed from betting markets.

¹⁸ The break point 2005 is chosen based on a graphical analysis of the structural changes of the soccer match effects over time.

¹⁹ For national teams, this variable is coded to be 1 if the market model includes a measure for a global market effect (e.g. MSCI world index is included in the examination of matches of the US national team on the S&P 500). For individual clubs, this variable is coded to be 1 if the market model includes a stock market index (e.g. FTSE 100 index is included in the examination of a UK soccer club’s stock price).

²⁰ The idea of the funnel asymmetry test follows from rotating the axes of the funnel plots in Figure (1) and inverting the values on the new horizontal axis (Stanley and Doucouliagos, 2012). A significant estimate of the slope coefficient then provides formal evidence for funnel asymmetry.

²¹ According to Doucouliagos and Stanley (2013), publication bias can be classified as ‘little to modest’ if the bias coefficient is statistically insignificant or $\left|{\hat{\beta }}_{SE}\right|<$ 1; ‘substantial’ if ${\hat{\beta }}_{SE}$ is statistically significant and 1 $\le \left|{\hat{\beta }}_{SE}\right|\le$ 2; and ‘severe’ if ${\hat{\beta }}_{SE}$ is statistically significant and $\left|{\hat{\beta }}_{SE}\right|>$ 2. These guidelines hold for the test of publication bias without including further moderator variables (Equation 2).

²² Bernile and Lyandres (2011) show that investors overestimate the winning probability by nearly 5 percentage points.