June 16, 2010

# Achieving the right consistency - I

A comprehensive statistical analysis of consistency for Test batsmen

 Mark Richardson wasn't the most attractive batsman, but with him you knew, more than with any other player, what you were going to get © Getty Images

My first few analyses for It Figures are all going to be broadly about the same thing, and that thing could broadly be called consistency. I'll bet that, at some time or other, everyone reading this post has criticised a cricketer for being inconsistent. I've done it myself but, whenever I have, I've had a nagging doubt: is performing brilliantly in one match and terribly in the next really any worse (or better) than being moderately good in two games on the trot? Maybe some stats can help us to unpick this issue.

I'm going to start by looking at batsmen. More specifically, my focus, in this first post, is batsmen's innings-to-innings consistency. If Batsman A has scores of 0, 138, 11, 0, & 101, and Batsman B has scores of 52, 50, 45, 48, & 55, then they both have the same average (50.00). However, there's a very obvious difference between the ways in which they've achieved the mark that we won't appreciate, if we concentrate on the average alone.

There are two big questions here, for me: (i) is it possible and instructive to identify batsmen with more or less consistent careers, and to quantify how much variability their records show? and (ii) does it matter? Is there any way in which a run of scores like Batsman A's is demonstrably better or worse - for himself and/or his team - than that of Batsman B?

Mister Hugely Reliable

S Rajesh comes close to answering the first of my questions in this It Figures avant la lettre column from 2006. He proposed a consistency index that is derived by dividing a batsman's average by the standard deviation (SD) of runs scored in each of his innings. I think he's on exactly the right lines, here, but I think the index can be improved in two ways. Firstly, I'm twitchy about combining one measure - the batting average - that makes an adjustment for not-out innings with another - the SD of the same dataset - that does not. For this reason, I'd rather rely on simple runs-per-innings (RPI), in this context. This way, both halves of the sum are quantifying the same thing and, although both may be affected by not-out innings, they are both affected equally. The second modification I have made is to turn the sum upside-down, so we have SD divided by RPI. Mathematically, this makes no difference to the ranking of results (although it means that low numbers, rather than high ones, indicate greater consistency).

The advantage of doing these two things is that the number you end up with has a solid interpretation: it is the percentage of deviation around the mean that is observed, on average, throughout the dataset. Dividing the SD by the mean is a trick statisticians use quite often; they call the result the coefficient of variation (CoV). As Rajesh pointed out, it's important to perform this scaling, rather than concentrating on SDs on their own, otherwise the batsmen who score most runs will always appear to have more variability in their records. A batsman with scores of 5, 30, and 100 has the same CoV as one with scores of 10, 60, and 200, though they have very different SDs.

So much for the theory; what about the results? Table 1 shows the batsmen who have been most and least consistent on an innings-to-innings basis throughout Test history, with a few notable figures picked out from the middle of the table.

Top of the lot is Kiwi opener Mark Richardson. He may not have set the world alight compared to some of his dashing contemporaries, but his solidity as an opening batsman can easily be overlooked: he reached double figures in 80% of his Test innings (a very high proportion, as noted in another Numbers Game a few years ago), and only ever registered one duck. What stopped him from threatening the real top rank of the game was that, though he'd seldom get out cheaply, he was also pretty unlikely to score very heavily, as a total of four centuries from 65 innings and a top score of 145 attests. These characteristics are perfect for a low CoV, because they imply that a large majority of his innings fell in a relatively tight range in the middle of possible scores. Cricket will always find a way of surprising you but, to a greater extent than with any other batsman, you knew what you were going to get from Richardson.

 Name M I R Ave RPI SD CoV 1. MH Richardson 38 65 2,776 44.77 42.71 35.16 0.823 2. H Sutcliffe 54 84 4,555 60.73 54.23 45.24 0.834 3. TL Goddard 41 78 2,516 34.47 32.26 27.11 0.840 4. SM Katich 49 85 3,792 48.00 44.61 38.72 0.868 5. MS Dhoni 43 66 2,428 42.60 36.79 32.36 0.880 6. JB Hobbs 60 102 5,410 56.95 53.04 46.68 0.880 7. IR Redpath 66 120 4,737 43.46 39.48 34.89 0.884 8. JB Stollmeyer 32 56 2,159 42.33 38.55 34.20 0.887 9. PE Richardson 34 56 2,061 37.47 36.80 32.86 0.893 10. A Ranatunga 93 155 5,105 35.70 32.94 29.44 0.894 ... 32. JH Kallis 136 229 10,760 54.62 46.99 44.95 0.957 ... 35. AR Border 156 265 11,174 50.56 42.17 40.49 0.960 ... 47. KP Pietersen 62 111 5,166 49.20 46.54 45.54 0.978 ... 56. DG Bradman 50 80 6,996 99.94 87.45 86.65 0.991 ... 85. RS Dravid 138 238 11,372 54.15 47.78 48.91 1.024 ... 97. RT Ponting 143 241 11,828 55.27 49.08 50.89 1.037 ... 107. Inzamam-ul-Haq 119 198 8,829 50.16 44.59 46.63 1.046 108. SR Tendulkar 166 271 13,447 55.57 49.62 51.92 1.046 ... 113. IVA Richards 121 182 8,540 50.24 46.92 49.43 1.053 ... 115. SM Gavaskar 124 214 10,122 51.12 47.30 49.96 1.056 116. SR Waugh 168 260 10,927 51.06 42.03 44.51 1.059 ... 131. GS Sobers 93 160 8,032 57.78 50.20 54.02 1.076 ... 226. BC Lara 130 230 11,912 53.18 51.79 62.43 1.205 ... 238. V Sehwag 75 128 6,608 53.72 51.63 64.71 1.254 ... 245. DW Randall 47 79 2,470 33.38 31.27 40.90 1.308 246. Zaheer Abbas 78 124 5,062 44.80 40.82 54.00 1.323 247. SE Gregory 58 100 2,282 24.54 22.82 30.33 1.329 248. LG Rowe 29 49 2,047 43.55 41.78 55.98 1.340 249. GJ Whittall 46 82 2,207 29.43 26.91 36.45 1.354 250. DL Amiss 50 88 3,612 46.31 41.05 55.74 1.358 251. MS Atapattu 90 156 5,502 39.02 35.27 49.93 1.416 252. Mohammad Ashraful 55 107 2,306 22.39 21.55 30.70 1.425 253. Wasim Akram 104 147 2,898 22.64 19.71 28.15 1.428 254. MH Mankad 44 72 2,109 31.48 29.29 46.06 1.572 qual. 2,000 Test runs; complete list available here

I was slightly surprised to see MS Dhoni riding high in this list. His reputation is for a more free-spirited kind of play than might be expected to generate a low CoV. But it turns out that any such assumptions do him a bit of a disservice: his Test record is that of a reliable runscorer, rather than a hit-or-miss gunslinger. Simon Katich's presence next to him is perhaps more in keeping with his reputation.

It is intriguing to see both Herbert Sutcliffe and Sir Jack Hobbs in the top half-dozen of this list. There could surely be no firmer foundation for a partnership as successful as theirs than the kind of shared dependability this statistic suggests. If they both had more mercurial profiles then, though they each might have scored as many runs, they would have been unlikely to have shared so many significant partnerships.

The fact that Jacques Kallis has fallen down the list somewhat compared to Rajesh's analysis is, to a small extent, a reflection of my slightly different methods, but it's more to do with the fact that his record has become a wee bit more inconsistent in the 4 years since Rajesh wrote his column.

According to this analysis, the least consistent batsman in Test history is Vinoo Mankad. His career has the opposite profile to Mark Richardson's: there is a very high proportion of low scores in his record (he only got into double figures 57% of the time) but, when he got in, he often went on to score big hundreds (including two doubles in one series against New Zealand in 1955/56). In contrast to Richardson's reliable-but-unspectacular record, Mankad's performances were an awful lot less predictable.

Wasim Akram's position at the bottom of the list is very largely ascribable to the effect of one mammoth score of 257* in the midst of a dataset that characteristically reflects a much more modest level of achievement (there's a good argument for calling this the most out-of-character innings in Test history, as discussed in a recent Ask Steven). If that one innings is excluded from his record, his CoV reverts to a much more run-of-the-mill 1.119.

Marvan Atapattu's status is probably not surprising for a man who started his Test career with a famous string of failures, but ended up with 6 double-centuries under his belt.

So...?

The unanswered question I find most intriguing is whether, in the grand scheme of things, any of this matters. As cricket fans, we're quite used to berating inconsistent batsmen ("you never know what you're going to get: one day, he's brilliant; the next, he couldn't buy a run") but, then again, we may have a paradoxical tendency to look down our noses at those with the least variable records ("he's good at getting in, but he never goes on to register a matchwinning score"). Is either of these positions more justifiable than the other?

I've come up with two ways of answering this question. The first is to examine whether consistent batsmen, ultimately, score more runs than their more mercurial counterparts. It's all very well to invent hypothetical 50-averaging batsmen with consistent and inconsistent records, like I did in my introduction, but it may be that, in the real cricketing world, batsmen with one profile or the other are more likely to achieve a decent average.

To explore this, I used a statistical technique called regression (to be more precise: univariate ordinary least squares linear regression), which enables us to assess the relationship between two variables. The results are shown in Figure 1. Each batsman's CoV is plotted against his average, with the typical relationship between the two (the regression line) indicated by the red dotted line. You can see that, although there's an awful lot of scatter around the trend, the datapoints generally appear to line up with a slight downwards slope. This suggests that there is a weak but identifiable association between the two variables, with more consistent batsmen tending to average slightly more (for any statsheads, that means that r 2 is a pretty dismal 0.065, but p<0.001 for the slope coefficient).

 Fig 1 Association between consistency (coefficient of variation) and success (average) for Test batsmen © Gabriel Rogers

Clearly, there are plenty of examples that do not fit the general trend too well, but it appears that, on average, consistency is associated with higher runscoring. Actually, a more pronounced correlation would have been surprising, because we didn't see a very obvious hierarchy in the consistency list - no one is suggesting that Mark Richardson was, in any meaningful way, a better batsman than Brian Lara. Nevertheless, it does seem to be the case that consistency is, by and large, a positive thing for individual batsmen. This may seem like an obvious finding, but I don't think it's been demonstrated before.

My second way to assess the value of batting consistency was to see whether it has a positive effect for the team. So I looked to see if there's any correlation between each batsman's CoV and his record of winning matches. I did this in exactly the same way, plotting one variable against the other, and drawing a univariate regression line through the results. For Test match cricket, there was a very weak, but still detectable, association between CoV and percentage of matches won (r 2=0.015; p=0.005); this vaguely suggests that, the more consistent a batsman is, the more likely he is to be on the winning side. It's a pretty unsatisfactory analysis, though, with an awful lot of noise around the hint of a signal. What I was more interested to find is that the correlation gets quite a bit stronger when, instead of winning record, you look at each batsman's not-losing record. The results of this analysis are shown in Figure 2. You can see a relatively shallow, but pretty obvious, upwards slope to the dataset, showing that, on average, the most consistent batsmen are also those who have lost the lowest proportion of the Test matches in which they have played.

 Fig 2 Association between consistency (coefficient of variation) and losing record for Test batsmen © Gabriel Rogers

The fact that consistency is associated with not-losing more strongly than it is with winning suggests that consistent batsmen really come into their own when it comes to securing draws for their teams. (And, indeed, regressing CoV against draw-rate produces a strongly significant result [p=0.002].) So, if you've got a team packed with consistent batsmen, you might not win too many more games, but you might draw some that less consistent teams would lose. I'm not quite sure how to explain this finding in cricketing terms; if you've got any bright ideas, please feel free to comment!

Once again from the top in pyjamas

The remainder of this post repeats the above analysis for ODI cricket.

Table 2 lists the most and least consistent batsmen in ODIs, The list is topped by Australia's two great "finishers" - Michaels Hussey and Bevan. We're used to seeing them high on lists of ODI stats, but it's worth remembering that - because CoV, as I have calculated it, relies on RPI rather than average - the high number of not-outs in each of their records has no direct influence on their excellent consistency ratings. Plenty of players have higher RPIs that these two; it's only once not-outs are factored in that their averages rise so high (although that doesn't necessarily mean the not-outs inflate their average, as is often assumed; Charles Davis has done good work on this). Accordingly, it is notable that consistency stats for these two players agree with their conventional records: I conclude that it was the dependability - as much as the volume - of their contributions that marked them out as matchwinners for their team.

 Name M I R Ave RPI SD CoV 1. MG Bevan 232 196 6,912 53.58 35.27 25.60 0.726 2. MEK Hussey 137 113 4,029 53.01 35.65 26.04 0.730 3. RR Sarwan 156 146 5,098 43.95 34.92 27.98 0.801 4. AH Jones 87 87 2,784 35.69 32.00 25.72 0.804 5. NH Fairbrother 75 71 2,092 39.47 29.46 23.81 0.808 6. IR Bell 78 76 2,483 35.47 32.67 26.45 0.810 7. GP Thorpe 82 77 2,380 37.19 30.91 25.04 0.810 8. CG Greenidge 128 127 5,134 45.04 40.43 33.01 0.817 9. AJ Lamb 122 118 4,010 39.31 33.98 27.87 0.820 10. RG Twose 87 81 2,717 38.81 33.54 27.54 0.821 11. Javed Miandad 233 218 7,381 41.70 33.86 27.81 0.821 12. DM Jones 164 161 6,068 44.62 37.69 31.03 0.823 ... 17. Zaheer Abbas 62 60 2,572 47.63 42.87 35.58 0.830 ... 20. ML Hayden 160 154 6,131 44.11 39.81 33.38 0.838 21. GC Smith 153 151 5,732 40.37 37.96 31.86 0.839 ... 24. S Chanderpaul 261 245 8,648 41.78 35.30 29.96 0.849 ... 26. Inzamam-ul-Haq 374 348 11,701 39.53 33.62 28.58 0.850 ... 28. MS Dhoni 159 141 5,246 50.44 37.21 31.79 0.855 29. JH Kallis 298 284 10,809 46.59 38.06 32.68 0.859 ... 31. Mohammad Yousuf 275 261 9,458 42.80 36.24 31.25 0.862 32. RS Dravid 335 309 10,644 39.42 34.45 29.71 0.863 ... 40. KP Pietersen 96 86 3,202 45.10 37.23 32.83 0.882 ... 49. RT Ponting 341 332 12,623 42.79 38.02 34.11 0.897 ... 55. IVA Richards 187 167 6,721 47.00 40.25 36.45 0.906 ... 74. BC Lara 295 285 10,348 40.90 36.31 34.28 0.944 ... 105. SR Tendulkar 442 431 17,598 45.12 40.83 40.14 0.983 ... 147. BB McCullum 171 145 3,569 29.02 24.61 26.77 1.088 148. DI Gower 114 111 3,170 30.78 28.56 31.17 1.091 149. Mohammad Ashraful 157 151 3,298 23.90 21.84 24.09 1.103 150. Kapil Dev 225 198 3,783 23.79 19.11 21.10 1.105 151. WW Hinds 119 111 2,880 28.51 25.95 28.83 1.111 152. RS Kaluwitharana 189 181 3,711 22.22 20.50 22.79 1.112 153. VVS Laxman 86 83 2,338 30.76 28.17 31.66 1.124 154. L Vincent 102 99 2,413 27.11 24.37 28.05 1.151 155. H Masakadza 95 95 2,601 28.58 27.38 32.86 1.200 156. KO Otieno 89 87 2,016 23.44 23.17 28.16 1.215 qual. 2,000 ODI runs; complete list available here

It's surprising to see Ian Bell so high in the list (actually, it's kinda surprising to learn that Ian Bell has 2,000 ODI runs). His reputation may not suggest limited-overs strength but, if you look at his record, it's clear that he has very seldom failed completely in ODIs. Of course, there's a relative lack of dramatic successes, too, and we've already seen that these two characteristics tend to produce a low CoV.

Another unexpected finding is that Sachin Tendulkar's ODI record is not an especially consistent one. To an extent, this is just because his scores encompass such a wide range. Obviously, Tendulkar has a much higher proportion of big scores in his record than someone like Ian Bell. Less predictably, however, he also has a higher proportion of cheap dismissals, and it is the swing from one extreme to the other that produces a higher CoV. Tendulkar's customary position at the top of the ODI order may provide part of the explanation for both of these features: if he hits his stride in any given innings, he will have full opportunity to score plenty of runs, and this may not be true of those who bat lower in the order; on the other hand, he also has the challenge of facing the new ball with close catchers in place - something that may raise his probability of early dismissal. It is notable that there are relatively few opening batsmen amongst those with the lowest CoVs (although I checked, and this is not a systematic bias in the dataset: batting position is not, in itself, predictive of consistency).

In contrast, it is anything but a shock to see Tendulkar's teammate, VVS Laxman, near the foot of the table: has any batsman more dramatically personified sublime-to-the-ridiculous swings of achievement in the ODI era?

I don't know if you've noticed it, but I think there's quite a difference between the ODI consistency list and the Test match equivalent, in Table 1, above. It seems to me that - the odd diminutive ginger anomaly aside - there is some sort of hierarchy going on in the ODI list. Players in the top part of the table are, by and large, better than those who come lower down. So we might expect to see a more pronounced correlation between ODI CoV and batting average than was the case in the analogous Test analysis. Any such expectation would be right on the money, as Figure 3 shows. There is a fairly strong association between the two variables (r 2=0.196; p<0.001), with a clear downward trend, suggesting that lower CoVs - indicating greater consistency - are associated with higher averages. I'm pretty sure that there is no mathematical reason why consistency should appear to be more valuable in ODI cricket than it is in the five-day game. If anyone can think of a cricketing reason, please do put it in the comments.

 Fig 3 Association between consistency (coefficient of variation) and success (average) for ODI batsmen © Gabriel Rogers

And, if it's positive for individual batsmen to be more consistent in their ODI runscoring, you'd expect their teams to see some benefit. As Figure 4 shows, this would appear to be another reasonable inference: batsmen with lower CoVs are, on average, those who win most ODIs (r 2=0.163; p<0.001).

 Fig 4 Association between consistency (coefficient of variation) and winning record for ODI batsmen © Gabriel Rogers

Conclusions

So what if, playground style, you're offered first pick between two batsmen with different records? Given a choice between a consistent batsman and an inconsistent one who averages more, you'd be a fool not to go for the one with the higher average. However, if your choice was between two batsmen with similar averages but different CoVs, I'd go for the more consistent one every time, as a result of what I've learned in this analysis. My expectation would be that he'd help me win - or at least not lose - a greater proportion of my matches. What is more, if I was provided with no information at all about the batsmen's averages, but did know their CoVs, I'd favour whoever had the more consistent record, because it would be a reasonable - though far from infallible - guess that he'd also have the better average.

One corollary of this conclusion may be that "matchwinning" performers are a bit of a myth. In the future, I want to do some work on whether there is such a thing as a true matchwinner in cricket (what analysts in other sports sometimes refer to as clutch players) but, with this analysis as my starting point, my provisional view is that the kind of batsman who quietly gets on with contributing on a match-to-match basis may be of at least as much value as one who has an exceptional game once in a while.

I've got some related posts coming up about consistency among bowlers, and swings of form over longer periods.

All stats calculated Jun 10, 2010 (i.e. all Tests up to England v Bangladesh at Manchester, Jun 4-6, 2010 [Test # 1959] and all ODIs up to Zimbabwe v Sri Lanka at Harare, Jun 9, 2010 [ODI # 2990]).

Technical appendix

Anyone who isn't incredibly fascinated by statistical methods doesn't need to read this bit, but I like to give a precise account of what I've done, in case anyone cares.

Technical note #1. My SDs are calculated as population SDs (i.e. if we're going to be really geeky, I have not adopted Bessel's correction). The reason for this is that, in batsmen's complete careers, we're dealing with all the observations that are available to us. This is unusual, for a statistician: normally, we have a limited sample of observations from which we want to draw inferences about a wider population (ask 1,000 people how they intend to vote, and you can predict what the entire electorate is going to do; give 500 people a drug, and you can tell how effective it'll be for everyone... that sort of thing). Here, though, the data we have is all we're going to get, so it's appropriate not to use the tiny correction that's normally strictly necessary. If anyone wants to replicate my analyses in Excel or Access, you need to use the StDevP function, not the normal StDev. If anyone else read the foregoing, didn't understand a word of it, but wonders whether it makes a difference to my outputs, the answer is no: the effect is tiny, but I believe it's more correct, so that's what I've done.

Technical note #2. Any statisticians reading this analysis might have been slightly concerned that the regression I presented in Figure 1 is unduly influenced by the highest-averaging batsmen (a phenomenon statisticians refer to as leverage). I did some sensitivity analyses that established that this isn't the case (p remains <0.001 when the dataset is restricted to batsmen averaging <60 or <50).

Technical note #3. Another concern statisticians might have with my regressions is that I've picked two covariates of consistency and analysed them separately (univariately). A more comprehensive model would be a multivariate one - that is, one that bundles everything up in the same analysis. So I did that. For Test cricket, I regressed CoV against average, losing percentage, and an interaction term. The only significant covariate was average (p<0.001). This suggests that the reason more consistent batsmen lose fewer Test matches is that they average more: there's no independent effect of consistency on not-losing. For ODIs, the multivariate results - regressing CoV against average, winning percentage, and an interaction term - are more interesting: all three covariates come up p<0.05 (and r 2 rises to 0.415). This suggests that more consistent batsmen are likely to win more games even if they don't average any higher, and the significant interaction term indicates that, the more games you win, the more consistency is of value in raising your average.

Technical note #4. The super eagle-eyed may have noticed that the scatterplots showing CoV -v- Average are rather bushier than those showing CoV -v- Results. This is because winning (or losing) percentages are constrained at both ends, and I found that the good number of players have won or lost all of their games were skewing results around, somewhat. Accordingly, all the CoV -v- Results analyses are limited to players with 40 or more innings. As a rule, I don't like doing this because, although tiny samples can produce weird results, their weirdness should balance out on either side of the average, so it's my preference to use all the data that's going. In this instance, though, I found I got much more sensible results by adopting an artificial constraint.

Technical note #5. I take the view that Australia v. ICC World XI, 14–17 October 2005, was not a Test match; similarly, these games are never included in my ODI stats.