feature

Read all about it?If statistics are of little help in predicting individual football results, are the professional followers of the sport any better? David Forrest and Robert Simmons analyse the performance of newspaper tipsters in helping the gambler to beat the bookies.

Can sports tipsters help? Evidence for English football

We live in a world where much complex in-formation relevant to forecasting the future is available from many sources, one of which is the print media. Newspapers give specialist advice on buying houses, cars and furniture, and on fi -nance. Th ey also give much information about football. Here a prime motive for readers want-ing probabilistic forecasts of future events is to give them an edge in the betting on match re-sults. We look at the eff ectiveness of tipster ad-vice published in daily newspapers and compare it with the performance of forecasts derived from formal statistical modelling, and from bookmakers’ odds themselves. We show that newspaper advice is rather ineff ective and that statistical modelling (of the kind outlined in the previous article) is a superior source of guid-ance. But it is bookmaker betting odds that are the most effi cient predictors of match results. Th is is perhaps unsurprising: newspaper tipster advice, for example, is free, whereas bookmakers would lose substantial sums of money if their odds were poor predictors of match results.

Tipsters

Until recently many daily newspapers off ered a column in their sports sections, usually on Monday or Tuesday, where a tipster would list a set of predictions of results for fi xtures on the upcoming weekend. In the Daily Mirror the tipster would off er a set of accompanying statistical tables as form guides, rather like the guides produced to aid horse race bettors. He might also off er some analytical commentary. Th e service was intended to off er free assist-ance to punters who took part in the football pools. Presumably editors felt that without

rate than following any of the tipsters. Th is is a consequence of the importance of “home advan-tage” that is universal across team sports. But the implied criticism here is unfair to the tipsters. Recall that their columns were designed to help bettors predict draws in the football pools game. Th e tipsters felt obliged to call at least some matches as draws in their predictions. Also, a game within the football pools enabled punters to identify a number of away wins so these also had to feature in tipster selections.

Ordered logit modelling

A more rigorous assessment of tipster perform-ance is by ordered logit regression analysis. In this method we think of teams as ordered (ranked) by relative strengths. We do not observe “team strength” as this includes such intangible factors as leadership and “team spirit”. Th ese relative strengths map into match outcomes through the logistic distribution. Th e dependent variable is then constructed as categorical, for example: a home win = 0, a draw = 1 and an away win = 2. Tipster forecasts can be treated as dummy vari-ables, for example corresponding to draw and away predictions with home win prediction as an omitted base category.

We can use the ordered logit regression model to answer three questions.

1. Do tipsters’ selections signifi cantly improve upon random selection?

2. Do tipsters’ selections off er something extra (unspecifi ed) to easily available public information?

3. Does a consensus forecast taken from the three tipsters outperform any single tipster?

To answer the fi rst question we regress match outcome on tipster dummy variables using the

this service fewer readers would buy their par-ticular newspapers.

Th is service has been replaced more re-cently by more specifi c advice—particularly in those papers which we still refer to as broad-sheets—where expert tipsters nominate “best bets” over a variety of sports events, which might be as diverse as a golf championship or American football games.

Evaluation of tipster performance

We thought it would be useful to establish sta-tistically whether free football tipster services were eff ective. Although the scope of this in-quiry was narrow, the methodology could use-fully be applied to other settings where special-ist advice is sought and is available at low cost.

In a detailed analysis published in the In-ternational Journal of Forecasting (2000)1 we looked at a set of 1694 English professional league football matches played between De-cember 1996 and April 1998. We inspected the match predictions of tipsters of three English newspapers: the Daily Mail, the Daily Mirror and Th e Times. Th ese were selected to give a representative cross-section of British newspa-per types and appeal over the population. Th e proportions of matches called successfully by the tipsters were:

Daily Mail 42.6%Daily Mirror 41.1%Th e Times 42.9%

Th e frequencies of actual match outcomes were:

Home wins 47.5%Draws 28.1%Away wins 24.4%

It immediately follows that always predicting a home win would have achieved a higher success

20 march2006

020 Forrest.indd 20020 Forrest.indd 20 09/03/2006 12:23:1009/03/2006 12:23:10

21march2006

ordered logit model and then compare maxim-ised log-likelihood for the unrestricted model with tipster dummies (LLU) with the maxim-ised log-likelihood for the restricted model containing just a constant, and of course an error term, LLR. Formally, the coeffi cients on tipster dummies in the restricted model are constrained equal to zero. In a conventional ordinary least squares regression model we have something analogous to an F-test on the joint signifi cance of regressors. Here, our test statistic for a likelihood ratio test follows a chi-squared distribution and is calculated as 2(LLU—LLR). With just two restrictions the critical value of the chi-squared statistic is 5.99. We fi nd that this value is easily exceeded for all three tipsters. So tipsters did improve on random selection in making their match predictions.

Th e second question is relevant because bettors may take two things from a tipster service, whether it is free or paid for. First, they may fi nd that a tipster can process available information more cheaply and more effi ciently than the individual can. Second, the tipster may have some additional knowledge that is not present in public information. In football this might be information about the form or habits of particular players. We can test for whether tipsters add additional private infor-mation by setting up some indicators of public information as regressors in the ordered logit model. We then ask whether tipster dummy variables add signifi cant explanatory power to the model over and above these public informa-tion variables. Th e public information variables included diff erences in league standings, ratios of goals scored to conceded for each team in a match using cumulative totals for the season thus far, and recent form over the previous fi ve games. Football match results are notoriously noisy and diffi cult to predict and we found that some of our public information variables did not signifi cantly add to the explanatory power of the model. Th is was true of recent form, for example. But some of our public in-formation variables were signifi cant predictors of match results. Using the likelihood ratio test described above, our restricted model had pub-lic information variables only, whereas the un-restricted model added tipster predictions. We found that there was only one tipster, the Daily Mail, for whom the addition of tipster predic-tions added signifi cant explanatory power to the model. Th e other two tipsters failed this particular test. Actually, in a further test, we found that tipsters did not even appear to make full use of the public information avail-able as summarised in our measures.

To answer the third question, we defi ned a consensus forecast as one where two or more tipsters agreed on a particular match predic-tion. We had three consensus dummy variables: home tip, draw tip and away tip. Again we as-sessed the performance of a consensus forecast using the likelihood ratio test. We found that the addition of consensus forecasts gave a test statistic of 10.68, to be compared with a criti-cal value for three degrees of freedom of 7.82. So this tells us that bettors could have made signifi cantly superior match predictions using consensus forecasting by combining tips from three sources. But, on the whole, our fi ndings cast serious doubt over whether newspaper tips on sporting outcomes are ever worth following.

Professors B. Boulier and H. Stekler, in an article in the International Journal of Fore-casting (2003)2, examined the eff ectiveness of the high-profi le sports editor of Th e New York

Times in predicting the points diff erence be-tween teams in American football games. Al-though the methodology employed diff ered from ours, the fi ndings were similar. Certainly the forecasts were not as accurate as bookmak-er spreads for forecasting outcomes, and so fol-lowing the expert could not yield betting gain. A novel feature of this article was that the au-thors also modelled how the expert responded to information like recent match scores. In-terestingly, the set of forecasts that the expert would have been predicted to make on the basis of modelling his typical behaviour was more eff ective than the set of actual forecasts. In other words, when the expert made subjec-tive adjustments to forecasts from his implicit model, this actually worsened performance. Th is echoes general fi ndings beyond sport that statistical modelling outperforms subjective forecasting in predicting future events. Per-haps experts typically give incorrect weighting to non-statistical information.

But in the article by Boulier and Stekler it is the betting market that is found to yield the most accurate set of forecasts for US foot-ball. Th e terms of bets off ered already display enough fi nesse that it is hard to beat them by consulting experts. Th is is unsurprising. Book-makers lose considerable sums of money if they set “incorrect” odds or spreads and so they may be expected themselves to recruit the best expertise.

In an article in the International Journal of Forecasting (2005)3 we returned to Eng-lish football to test the effi ciency of odds set by bookmakers. We pitted them, over fi ve seasons (1998–2003), against a set of proba-bilistic forecasts generated by a sophisticated, information-rich statistical model, constructed by our collaborator Professor John Goddard, using data from 15 preceding seasons. Th e formal framework was to compare forecasts derived from the odds alone and those from the statistical model. Th e statistical model fared much better than the newspaper tipsters had in our earlier analysis. Indeed, in the early seasons it performed a little better than the odds and could have been the basis of a betting strategy that would have permitted bettors to just about break even rather than sustain the customary losses. However, odds proved an increasingly eff ective forecasting tool and by the end of the period dominated the statisti-cal model. Using the model output to guide betting would, in the later years, simply have been a way of losing money. Th is trend to-wards computer-generated forecasts failing to identify good value bets is consistent with bookmakers demonstrating more skill as their margins fell under pressure of Internet compe-tition. With smaller margins, the penalties of posting inaccurate odds increase.

So it appears harder than ever to beat the bookmaker, whether you consult subjective newspaper tipsters or statistical experts. Th e problem for the bettor is that odds setters have a very special skill. In fact, it is so special that odds setters are typically paid salaries double those of full professors of statistics in UK uni-versities!

References1. Forrest, D. and Simmons, R. (2000) Fore-

casting sport: the behaviour and performance of football tipsters. International Journal of Forecasting, 16, 317–331.

2. Boulier, B. and Stekler, H. (2003) Pre-dicting the outcomes of National Football League games. International Journal of Forecasting, 19, 257–270.

3. Forrest, D., Simmons, R. and Goddard, J. (2005) Odds setters as forecasting: the case of the football betting market. International Journal of Forecasting, 21, 551–564.

Dr David Forrest is Reader in Economics at the Uni-versity of Salford, specialising in sport economics and the economics of gambling, and with Dr Robert Sim-mons has co-authored numerous papers on betting in sport. Dr Simmons, Senior Lecturer at Lancaster University, has an international reputation as a sports economist. He is also a football referee.

“There is serious doubt over whether newspaper tips on sporting events are ever

worth following”

020 Forrest.indd 21020 Forrest.indd 21 09/03/2006 12:23:1209/03/2006 12:23:12