In a fluid betting market, odds will change in response to the weight of money on the different possible outcomes for that market. If one outcome sees more interest than another, its price will shorten, whilst the other(s) will lengthen. A shortening price is called a steamer, and lengthening one, a drifter.
When a betting market opens, earlier prices reflect only a limited number of opinions about the possible outcomes for that market. Consequently, these earlier prices are less likely to accurately reflect the “true” probabilities of each outcome, and as such the market is more likely to be informationally inefficient. The price adjustments that follow should, in theory, force the market towards a higher degree of efficiency, as more money representing an ever increasing number of opinions flows in. A priori we might reasonably expect the highest level of efficiency, on average, to be achieved as the market closes, being as it is now built on the largest spread of market opinions, and money wagered on them. If that is the case this would imply that those betting propositions that had steamed towards market closure may have at one time offered value, although of course such knowledge can only be fully acquired retrospectively. To test this hypothesis, and to analyse whether any edge can be found by understanding how a market moves prior to its closure, I have analysed two full seasons of professional English league football. Arguably, this sample size is relatively small. Nevertheless, it is a start. This article discusses the results and conclusions of this analysis, and specifically whether there is any value to be found following steamers.
Jack Houghton, writing in an article in 2008 on steamers and drifters for Betfair, did not have much time for this hypothesis. Analysing a sample of 65,000 horses running in 2006, he noted the prices 2 hours before the off and compared them to their starting prices. Then, categorising steamers as those runners whose odds had shortened by at least 5% and drifters whose odds had lengthened by at least the same amount, he found that from £10 level staking, the steamers would have lost nearly £3,500 whilst the drifters would have won nearly £2,310. Unfortunately Jack hasn’t made it clear what proportion of the 65,000 runners were steamers and drifters so we are unable to calculate yields. Furthermore, he doesn’t make it clear which price has been used to calculate those returns, the starting price or the price 2 hours before the off. Clearly, without a time machine we are not in a position to know whether every price will continue to steam for the next two hours, so calculating the returns using the original prices would essentially provide just a hypothetical result. Yet returning a loss to starting prices from steamers does not necessarily imply that we wouldn’t return a profit from the pre-steamed betting odds (assuming we could have known which all the steamers were 2 hours before the off). Again, since Jack has provided no details about actual prices, either starting or those 2 hours prior, we are not able to shed any light on this, and specifically determine whether there is any significant difference between these two sets of betting returns. All we can say in this case is that, and aside from the influence of chance, if these returns were based on starting prices, punters might well haven have over-bet the steamers and shortened their prices too much, whilst under-betting the drifters, forcing a lengthening of prices that went beyond what was appropriate, based on the actual results of the races.
That was horse racing in 2006; what about steamers and drifters for sports betting in more recent years? To get a handle on this I decided to compare closing market average bookmaker prices for all professional English league football matches played during the 2010/11 and 2011/12 seasons to average prices collected in advance of start times. Whilst each season contains a total of 2,036 matches for a combined total of 4,072, the actual data set analysed contained 4,069 matches for a total of 12,207 betting selections. For the 2010/11 season betting odds for 1 match were missing, whilst for a further 2 matches the average odds calculation was based on fewer than 10 bookmakers, and hence were omitted. The closing prices were provided by the odds comparison oddsportal.com, whilst the pre-closing prices came from the odds comparison betbrain.com. A couple of obvious limitations exist for this analysis. Firstly, I am using two different data sources, clearly not ideal, given the slightly different make-up of bookmakers used to compute the average betting prices. Secondly, there was no fixed time period for the duration between the original price collected and its closing price. With data from Betbrain collected on Friday afternoons, a Friday evening match started within a few hours of the time the odds were collected whilst, for the remainder of the weekend games, a larger time period existed between the Betbrain snapshot prices and the Oddsportal closing prices. Similarly, for midweek games, with Betbrain odds generally collected on Tuesdays mornings, and matches played on either Tuesday or Wednesday evenings, again the time period over which prices could shift was variable. However, without access to better quality data, this was the best I could do. In the event, the average price of all home wins, draws and away wins for the Betbrain and Oddsportal data samples was very close, 3.186 and 3.195 respectively, and certainly not significantly different suggesting that at least the use of two different data sources had probably not influenced the results in any meaningful way.
For this English football league match data set, steamers and drifters were categorised according to the percentage change in win expectancy defined by the betting odds. The largest increase in win expectancy was 10.7%, whilst the largest decrease was 10.4%. Only 1% of odds movements saw a greater than 5% shift in the win expectancy, either higher or lower. This is probably far smaller than in horse racing where the large numbers of runners per race and the greater amount of uncertainty inherent in the generally higher odds markets will create larger swings of opinion, and hence larger movements in betting price. It should be remembered, however, that these football odds data represent averages; presumably there would be bigger price swings if individual bookmaker prices were considered. Nevertheless, the results from the analysis tabulated below are revealing to say the least.
Theoretical achievable returns to average betting prices from steamers and drifters in the 2010/11 and 2011/12 English football league seasons
|Win expectancy change (WEC)||Bets||Average pre-closing odds||Average closing odds||Yield (pre-closing odds)||Yield (closing odds)|
|WEC > +2%||962||2.56||2.34||6.74%||-1.49%|
|1% < WEC ≤ +2%||1,195||2.85||2.72||3.83%||-0.29%|
|0% < WEC ≤ 1%||3,573||3.32||3.27||-5.08%||-6.39%|
|-1% < WEC ≤ 0%||4,518||3.35||3.39||-9.25%||-8.31%|
|-2% < WEC ≤ -1%||1,160||3.25||3.43||-19.16%||-15.51%|
|WEC ≤ -2%||799||2.83||3.13||-18.18%||-10.30%|
Blindly backing every steamer at the archived pre-closing prices would have lost the bettor just 1.24% on turnover. By contrast, backing every drifter (including those prices that saw no price movement at all) would have lost him over 12%. This difference is highly significant (p-value = 0.00002). Furthermore, on the 962 occasions where the win expectancy increased by more than 2%, he would have actually returned a profit of nearly 7% (provided of course he would have known that these selections would all steam by more than 2%). Compare that to the 799 win probabilities which drifted (or decreased) by 2% or more. In this case losses would be approaching a whopping -20%. The strength of this influence is staggering, indeed almost too large to be believed based on the magnitude of the price movements. Given the slightly shorter prices in general for steamers as opposed to drifters in this sample some of this effect may be a consequence of the favourite–longshot bias, but its influence is probably only marginal.
Of course these returns are purely theoretical since no-one could possibly predict all the betting prices that would go on to shorten by kick-off and all those which would lengthen. Nevertheless, the direction in which the market is moving at the time the odds were observed would surely provide a useful indicator. So let’s also look at the returns from betting closing prices. The influence of price shift is now understandably weaker. Indeed, if our original hypothesis – that the closing prices should provide the best approximation to fair prices – is true, then betting yields from closing prices should, in theory, be roughly equivalent regardless of whether prices had been steaming or drifting. Yet strangely we still find that returns from backing steamers are marginally better (-4.29%) than for blanket betting (-7.24%), whilst returns from backing drifters are marginally worse (-9.85%). This time the difference is not statistically significant at the 1% level (p-value = 0.017), although it is still apparent that backing the heavy steamers would barely lose the bettor any money at all, despite those prices having witnessed the most shortening prior to kick-off. Statistical significance aside, there remains the suggestion, contrary to what Jack Houghton had observed for horse races, that steamers still remain under-bet and marginally over-priced even at kick-off.
In conclusion, this analysis provides convincing support that the direction in which a win-draw-win football betting market is moving can be a useful indicator, even on its own, for result prediction. Indeed, it’s conceivable that a heavily steaming/drifting betting market remains inefficient, with residual value in the steamer, even at the time of market closure. It would certainly be intriguing to see if such a finding could be replicated for other seasons, other football leagues and other sports. Whilst it’s possible that this sample might be unrepresentative of a much wider population of football match betting odds data, it was the case when I tested this hypothesis for the first time I used only data from the 2010/11 season. Going back to analyse the data for 2011/12 once that season had been completed, the results were found to entirely replicate those from the earlier season.