Skip to main content
Log in

Delusions of Expertise: The High Standard of Proof Needed to Demonstrate Skills at Horserace Handicapping

Journal of Gambling Studies Aims and scope Submit manuscript

Abstract

Gamblers who participate in skill-oriented games (such as poker and sports-betting) are motivated to win over the long-term, and some monitor their betting outcomes to evaluate their performance and proficiency. In this study of Australian off-track horserace betting, we investigated which levels of sustained returns would be required to establish evidence of skill/expertise. We modelled a random strategy to simulate ‘naïve’ play, in which equal bets were placed on randomly selected horses using a representative sample of 211 weekend races. Results from a Monte Carlo simulation yielded a distribution of return-on-investments for varying number of bets (N), showing surprising volatility, even after a large number of repeated bets. After adjusting for the house advantage, a gambler would have to place over 10,000 bets in individual races with net returns exceeding 9 % to be reasonably considered an expert punter (α = .05). Moreover, a record of fewer bets would require even greater returns for demonstrating expertise. As such, validated expertise is likely to be rare among race bettors. We argue that the counter-intuitively high threshold for demonstrating expertise by tracking historical performance is likely to exacerbate known cognitive biases in self-evaluation of expertise.

This is a preview of subscription content, log in via an institution to check access.

Access this article

Price excludes VAT (USA)
Tax calculation will be finalised during checkout.

Instant access to the full article PDF.

Institutional subscriptions

Fig. 1
Fig. 2
Fig. 3
Fig. 4
Fig. 5

Similar content being viewed by others

Notes

  1. This bias against the player may take form of share-transaction fees, or the operator dividend incorporated into racing totalisator systems. By definition, at least 50 % (probably far more, since the distribution of expertise/knowledge tends to be positively skewed) of players participating in games with a skill component do not possess above-average knowledge or proficiency. Furthermore, above-average proficiency may not be sufficient to overcome the intrinsic bias against the player in the return schedule.

  2. Note that, by this definition, expert strategies may still yield long-term negative aggregate returns, if the benefit conferred by the expert strategy is outweighed by the intrinsic weighting of the game against the player.

  3. Equal prior probabilities assumes that non-expert players Other forms of random play may also be defined. Possibilities include using probabilities determined by starter number, starting odds, or by a particular player’s prior probabilities of with respect to starter or odds. These would generally result in a more sensitive benchmark for comparison (at the expense of increased complexity), so the impact of varying the random model might warrant further investigation.

  4. Once again, more complex random models may be defined that place variable bet sizes, potentially matched according to the distribution of bet sizes for the particular play record under consideration. For clarity, we omit such elaborations in the present paper.

References

  • Blaszczynski, A., & Nower, L. (2002). A pathways model of problem and pathological gambling. Addiction, 97(5), 487–499. doi:10.1046/j.1360-0443.2002.00015.x.

    Article  PubMed  Google Scholar 

  • Brownstein, A. L., Read, S. J., & Simon, D. (2004). Bias at the racetrack: Effects of individual expertise and task importance on predecision reevaluation of alternatives. Personality and Social Psychology Bulletin, 30(7), 891–904. doi:10.1177/0146167204264083.

    Article  PubMed  Google Scholar 

  • Chantal, Y., & Vallerand, R. J. (1996). Skill versus luck: A motivational analysis of gambling involvement. Journal of Gambling Studies, 12(4), 407–418. doi:10.1007/BF01539185.

    Article  CAS  PubMed  Google Scholar 

  • Dixon, M. J., Fugelsang, J. A., MacLaren, V. V., & Harrigan, K. A. (2013). Gamblers can discriminate “tight” from “loose” electronic gambling machines. International Gambling Studies, 13(1), 98–111. doi:10.1080/14459795.2012.712151.

    Article  Google Scholar 

  • Ehrlinger, J., & Dunning, D. (2003). How chronic self-views influence (and potentially mislead) estimates of performance. Journal of Personality and Social Psychology, 84(1), 5–17.

    Article  PubMed  Google Scholar 

  • Emond, M. S., & Marmurek, H. H. C. (2010). Gambling related cognitions mediate the association between thinking style and problem gambling severity. Journal of Gambling Studies, 26(2), 257–267. doi:10.1007/s10899-009-9164-6.

    Article  PubMed  Google Scholar 

  • Epstein, R. A. (2012). The theory of gambling and statistical logic. Waltham, MA: Academic Press.

    Google Scholar 

  • Garrett, T. A., & Sobel, R. S. (1999). Gamblers favor skewness, not risk: Further evidence from United States’ lottery games. Economics Letters, 63(1), 85–90. doi:10.1016/S0165-1765(99)00012-9.

    Article  Google Scholar 

  • Harper, J. D., & Ross, K. A. (2005). Stopping strategies and gambler’s ruin. Mathematics Magazine, 78(4), 255–268. doi:10.2307/30044172.

    Article  Google Scholar 

  • Klenke, A. (2008). Probability theory: A comprehensive course. New York: Springer.

    Book  Google Scholar 

  • Kruger, J., & Dunning, D. (1999). Unskilled and unaware of it: How difficulties in recognizing one’s own incompetence lead to inflated self-assessments. Journal of Personality and Social Psychology, 77(6), 1121–1134.

    Article  CAS  PubMed  Google Scholar 

  • May, P. (2013). In search of the winning system. Compton: Racing Post & Raceform.

    Google Scholar 

  • McMullan, J. L., & Kervin, M. (2012). Selling internet gambling: Advertising, new media and the content of poker promotion. International Journal of Mental Health and Addiction, 10(5), 622–645. doi:10.1007/s11469-011-9336-3.

    Article  Google Scholar 

  • Meadow, W., & Sunstein, C. R. (2001). Statistics, not experts. Duke Law Journal, 51(2), 629. doi:10.2307/1373203.

    Article  CAS  PubMed  Google Scholar 

  • Metcalfe, J. (1998). Cognitive optimism: Self-deception or memory-based processing heuristics? Personality and Social Psychology Review, 2(2), 100–110. doi:10.1207/s15327957pspr0202_3.

    Article  CAS  PubMed  Google Scholar 

  • Meyer, G., Meduna, M. von, Brosowski, T., & Hayer, T. Is poker a game of skill or chance? A quasi-experimental study. Journal of Gambling Studies, 1–16. doi:10.1007/s10899-012-9327-8.

  • Mitrovic, D. V., & Brown, J. (2009). Poker mania and problem gambling: A study of distorted cognitions, motivation and alexithymia. Journal of gambling studies/co-sponsored by the National Council on Problem Gambling and Institute for the Study of Gambling and Commercial Gaming, 25(4), 489–502. doi:10.1007/s10899-009-9140-1.

    Google Scholar 

  • Mun, J. (2006). Modeling risk: Applying Monte Carlo simulation, real options analysis, forecasting, and optimization techniques. Hoboken, NJ: Wiley.

    Google Scholar 

  • Neighbors, C., Lostutter, T. W., Cronce, J. M., & Larimer, M. E. (2002). Exploring college student gambling motivation. Journal of Gambling Studies, 18(4), 361–370. doi:10.1023/A:1021065116500.

    Article  PubMed Central  PubMed  Google Scholar 

  • Pelletier, M., & Ladouceur, R. (2007). The effect of knowledge of mathematics on gambling behaviours and erroneous perceptions. International Journal of Psychology, 42(2), 134–140. doi:10.1080/00207590600788047.

    Article  Google Scholar 

  • Pitt, M. K., Silva, R. S., Giordani, P., & Kohn, R. (2012). On some properties of Markov chain Monte Carlo simulation methods based on the particle filter. Journal of Econometrics, 171(2), 134–151. doi:10.1016/j.jeconom.2012.06.004.

    Article  Google Scholar 

  • Richard, D. C. S., Blaszczynski, A., & Nower, L. (2013). The Wiley-Blackwell handbook of disordered gambling. Hoboken, NJ: Wiley.

    Book  Google Scholar 

  • Schmidt, I. W., Berg, I. J., & Deelman, B. G. (1999). Illusory superiority in self-reported memory of older adults. Aging, Neuropsychology, and Cognition, 6(4), 288–301. doi:10.1076/1382-5585(199912)06:04;1-B;FT288.

    Article  Google Scholar 

  • Smith, M. A., & Vaughan Williams, L. (2010). Forecasting horse race outcomes: New evidence on odds bias in UK betting markets. International Journal of Forecasting, 26(3), 543–550. doi:10.1016/j.ijforecast.2009.12.014.

    Article  Google Scholar 

  • Team, R. D. C. (2011). R: A language and environment for statistical computing. Vienna, Austria. Retrieved from http://www.R-project.org/.

  • Toneatto, T., Blitz-Miller, T., Calderwood, K., Dragonetti, R., & Tsanos, A. (1997). Cognitive distortions in heavy gambling. Journal of Gambling Studies, 13(3), 253–266. doi:10.1023/A:1024983300428.

    Article  CAS  PubMed  Google Scholar 

  • Turner, N. (2013). Probability, random events and the mathematics of gambling. Problem Gambling Institute of Ontario. Retrieved from http://www.problemgambling.ca.

  • Tversky, A., & Kahneman, D. (1975). Judgment under uncertainty: Heuristics and biases. In D. Wendt & C. Vlek (Eds.), Utility, probability, and human decision making (pp. 141–162). Netherlands: Springer. Retrieved from http://link.springer.com/chapter/10.1007/978-94-010-1834-0_8.

  • Waylen, A. E., Horswill, M. S., Alexander, J. L., & McKenna, F. P. (2004). Do expert drivers have a reduced illusion of superiority? Transportation Research Part F: Traffic Psychology and Behaviour, 7(4–5), 323–331. doi:10.1016/j.trf.2004.09.009.

    Article  Google Scholar 

  • Williams, R. J., & Connolly, D. (2006). Does learning about the mathematics of gambling change gambling behavior? Psychology of Addictive Behaviors, 20(1), 62–68. doi:10.1037/0893-164X.20.1.62.

    Article  PubMed  Google Scholar 

Download references

Author information

Authors and Affiliations

Authors

Corresponding author

Correspondence to Matthew Browne.

Appendix

Appendix

Process Model

We define an expert strategy as one that generates long-term aggregate returns significantly greater than those generated by random play.Footnote 2 In horse racing, the simplest form of random play involves probable random horse selection in each horse race with equal probabilities, although other more complex forms of random play are possible.Footnote 3 Any actual aggregate return-on-investment (ROI) (and number of bets in which it was achieved) may be compared to the distribution of ROIs yielded by random play. The probability of observing the actual ROI is correctly interpreted as the probability of making an equal or greater return by random play. The null-hypothesis is that a given observed ROI is drawn from this population of random returns. Rejection of the null hypothesis entails that the observed return would be unlikely to have occurred by chance, and we may then accept the alternate conclusion that the observed strategy is expert.

In order to address this, we consider an imaginary ‘naïve’ player of horse racing, who has no expert knowledge, and who implements the simple strategy of placing a bet of $1 on a randomly selected horse (starter) on each race.Footnote 4 We assume only simple betting on a single horse to win, and do not consider more exotic forms of betting (e.g., win, place or show). Denote the race index as \(k \in \left\{ {1,2, \ldots ,K} \right\}\) and the number of starters in each race as N k . The selection of the nth starter (\(n \in \left\{ {1,2, \ldots ,N} \right\}\)) would entail a raw return for the player on that race of \(r_{k}^{n}\):

$$r_{k}^{n} = \left\{ {\begin{array}{*{20}c} {\left. {r_{k}^{*} } \right|n_{k}^{ + } = n_{k}^{*} } \\ { 0 otherwise} \\ \end{array} } \right.$$

where \(r_{k}^{*}\) is the payoff for selecting the winning horse \(n_{k}^{*}\) (determined by the totalisator or equivalent system), and \(n_{k}^{ + }\) is actual selection made. The expected return of a random player choosing a single starting horse per race to win is

$$E\left( {r_{k} } \right) = \mathop \sum \limits_{n} P\left( {r_{k}^{n} } \right)r_{k}^{n}$$

with \(P\left( {r_{k}^{n} } \right) = P\left( {n_{k}^{ + } = n} \right)\) being the probability of selecting each starter and \(\mathop \sum \nolimits P\left( {r_{k}^{n} } \right) = 1\). Since the naïve player selects horses with equal probability, they have a 1/N k probability of choosing the winning horse, other returns are zero, so the expected return is\(E\left( {r_{k} } \right) = N_{k}^{ - 1} r_{k}^{*}\).

Since the bet size is constant $1 for each race, the expected average ROI (R)

$$R^{K} = K^{ - 1} \mathop \sum \limits_{k} r_{k}$$

gives the ratio of total return to total investment over K races. Assuming the totalisator system allocates a standard proportion \(Q \in \left( {0,1} \right)\) of bets to winning players each race, and if the totalisator system were a perfect reflection of the underlying odds of winning, then

$$E\left( {R^{K} } \right) = E\left( {r_{k} } \right) = Q.$$

That is, the expected average return of a random player does not vary with the number of bets, and would be determined by the constant Q (currently approximately .85 in the Australian TAB). Note that this a prior expectation, given no knowledge of the odds of the payoff \(r_{k}^{*}\) of the horse the ultimately won. The posterior expectation \(E\left( {\left. {R^{K} } \right|r_{k}^{*} } \right) > Q\) (i.e. random players earn more) in races where the long-odds horse won, with the converse being true for races in which a short-odds horse won. If expert knowledge is not equally distributed in the population, then the totalisator system would not be a perfect reflection of the underlying odds of winning, and therefore \(E\left( {R^{K} } \right) < Q\), with the discrepancy in expected winnings being distributed to players with more expertise.

A human player picking starters on the basis of expert judgement has the hypothesis that their expertise will provide them with a probability of selecting the winning horse greater than the odds for that horse. We can summarise the effect of expert play as an ‘edge’ J > 0, that will improve their expected returns

$$E\left( {\widehat{R}^{k} } \right) = Q + J$$

which implies they may expect to experience positive returns only if J > 1 − Q, as \(K \to \infty\).

Analysis

A gambler deciding whether to persevere with their strategy in future (and indeed, whether to continue gambling at all) must first make the assumption that past return performance \(\left\{ {\widehat{r}_{1} ,\widehat{r}_{2} ,\widehat{r} \cdots \widehat{r}_{K} } \right\}, \widehat{R}^{K} = K^{ - 1} \sum {\widehat{r}_{K} }\) (using virtual or actual money) is a valid indicator of future performance. Assuming above average returns have been achieved, the next issue is to determine to what degree this play-history is a reliable indicator of underlying expertise (or an effective strategy) yielding: J > 0. Applying a standard frequentist perspective, we invert the problem by evaluating the null hypothesis: Assuming that J = 0, what is the probability of observing \(\widehat{R}^{K}\) (a player’s observed return after K bets) by chance alone? This is equivalent to determining the probability of drawing \(\widehat{R}^{K}\) or greater from the population of random returns \(R^{K}\).

Since modelling the distribution of the population of random returns \(R^{K}\) is critical to providing a rigorous basis for evaluating a given player’s return \(\widehat{R}^{K}\), it is worth considering its distributional properties. First, we note that it is parameterised broadly K and Q, but more exactly by the set of paired race parameters \(\left\{ {r_{k}^{*} ,N_{k} } \right\}KNN = 2N = 10R^{K}\), which determine the two point distribution of win (\(r_{k} = r_{k}^{*}\)) loss (\(r_{k} = 0\)) events for player picking a random starter in each race. In order to generalise, we must assume that each pair of race parameters in \(\left\{ {r_{k}^{*} ,N_{k} } \right\}\) are drawn from some joint distribution \(G\left( {r^{*} ,N} \right)\) with properties of independence and stationarity. Thus, \(\Pr \left( {R^{K} } \right) = F\left( {r^{*} ,N,K} \right)\), with \(G\left( {r^{*} ,N} \right)\) capturing the distribution of relevant race-characteristics: the number of starters and the spread of long-odds versus short-odds wins. It is not necessary that the data used to approximate \(G\left( {r^{*} ,N} \right)\) is the same as that that used to calculate \(\widehat{R}^{K}\) (i.e. \(\left\{ {r_{k}^{*} ,N_{k} } \right\}\)), as long as we can made the assumption that the distributional properties of \(\left\{ {r_{k}^{*} ,N_{k} } \right\}\) are similar to \(G\left( {r^{*} ,N} \right)\). In other words, we require the observed races to be a representative sample of the population of races of interest, in terms of their distribution of winners with respect to their starting odds. This provides the potential to generate standard reference tables for homogenous race populations (e.g., the Australian TAB), allowing practitioners to do hypothesis testing for a particular \(\widehat{R}^{K}\), without having to recreate the null distribution of returns. To combine some of the notation above, we restate the goal of the analysis as the following: calculate \(\Pr \left( {R^{K} > \widehat{R}^{K} } \right)\), assuming \(\widehat{R}^{K} \sim R^{K} = F\left( {r^{*} ,N,K} \right)\).

Classical theory does not provide much practical help in estimating F. The law of large numbers leads us to expect that \(R^{K}\) converges to \(E\left( {R^{K} } \right) = Q\) as \(K \to \infty\). However this tells us little regarding the dispersion or shape of observed returns around the mean of Q for realistic bet-run lengths K. From our description of the generating process, and assuming no prior information about the race, the distribution of \(R^{K = 1}\) is directly given by the distribution\(G\left( {r^{*} ,N} \right)\): it is a mixture, with majority probability concentrated at zero, with a very long-tail of non-zero returns following an inverse relationship with \(\Pr \left( {R^{K} } \right) \cong 1/R^{K}\). As K becomes very large, understanding the shape of F is also straight-forward. As we are considering aggregate return, and given the distribution of \(R^{K}\) is modelled as a sum of independent and identically distributed race returns, the central limit theorem states that the distribution of \(R^{K}\) converges to the normal distribution as \(K \to \infty .\)

For intermediate values of K, the shape of F may display significant deviations from normality. Even considering simplified approximations \(G\left( Q \right) \cong G\left( {r^{*} ,N} \right)\), deriving closed-form expressions for F(Q,K) does not appear to be a practical option, as the number of horses and distribution of odds varies randomly from race to race. We avoid such difficulties by simply adopting a non-parametric approach, and implement a Monte-Carlo sampling method for estimating the distribution of \(R^{K} \sim F\) for various bet-run-lengths K by repeatedly re-sampling a representative and homogenous sample of races with a constant Q expected return.

An efficient approach to sampling an empirical distribution of \(R^{K}\) from \(K = 1,2, \ldots ,K^{\hbox{max} }\) is via Monte Carlo simulation (Mun 2006; Pitt et al. 2012); i.e. to simulate a large set of J play histories, and to record the ROI after each bet, at each replicate: \(r_{j,K}\). Each history of raw returns \(\widehat{r}_{j,K}\) amounts to a Markov chain with transition probabilities determined by \(\left\{ {r_{m}^{*} ,N_{m} } \right\}\) (assumed to be representative of \(G\left( {r^{*} ,N} \right)\)). The distribution of progressive returns is used to generate statistics of interest, such as the .9, .95, and .99 ‰. Though moderately computationally intensive with contemporary computational resources, Algorithm 1 is conceptually very simple, and as noted above, may be pre-computed in reference tables for practical use in varying gambling contexts. Algorithm 1 is written in terms of the intuitive process of (a) selecting a random race from the ‘library’ of races, and then (b) selecting a random horse, at each step. However, programmers should be aware that it may be written and implemented most efficiently by bootstrap resampling from the full return distribution \(\left\{ {r_{m}^{n} } \right\}\) at each Markov transition.

All analyses and plotting was undertaken in the statistical programming environment ‘R’ (Team 2011). Code is available on request.

figure a

Rights and permissions

Reprints and permissions

About this article

Cite this article

Browne, M., Rockloff, M.J., Blaszcynski, A. et al. Delusions of Expertise: The High Standard of Proof Needed to Demonstrate Skills at Horserace Handicapping. J Gambl Stud 31, 73–89 (2015). https://doi.org/10.1007/s10899-013-9420-7

Download citation

  • Published:

  • Issue Date:

  • DOI: https://doi.org/10.1007/s10899-013-9420-7

Keywords

Navigation