11. A Revised Elo System

 

     

   

We now come to the crux of our argument, which focuses on the interpretation of the established rating, represented here in abstract form as the differential

[11.1]                       D R  =  d D P .

As Elo saw it, the established rating was essentially a small shift along the Percentage Expectancy Curve.  The derivative d is based on the slope of the curve, and the change in rating is calculated from the curve as the rating difference associated with the change in percentage score.  The interpretation offered here is quite different.  The change in percentage score arises from cumulative averaging as a correction of the estimate of long-term percentage score.  The correction is translated by the derivative d of the performance rating with respect to P into the rating change.  The different interpretation, as it happens, does not affect practical results, but it does tend to rule out the need for the Percentage Expectancy Curve.

We shall illustrate these different points of view by deriving a new version of the established formula, and by implication a new rating system.  The new system is logarithmic, like the Elo System, and uses the same constants.  The formula for percentage expectancy is also the same as the one derived by Elo,

[11.2]                Pe  =  1 / (1 + 10-D/400) ,

but not from a probability distribution.  To avoid problems with averaging, we shall rate results one game at a time.  The cumulative average for percentage score is consequently

[11.3]               D P  =  (S – Pe) / No .

To translate this into a rating change we take the derivative of the performance formula [5.6] derived in our discussion of ratio ratings,

[11.4]               R' =  400 / [P(1 – P) . ln10] .

The new established rating formula is therefore

[11.5]              D R  =  400(S – Pe) / [Pe(1 – Pe) . No . ln10]

where Pe is defined in [11.2].  The rating change resulting from a win in a pairing of various rating differences, positive and negative, is reported in the table below and compared with the rating change for the Elo System, interval and ratio.  The rating changes resulting from a loss are obtained by changing S as well as the sign of D, giving negative versions of the reported values.  The program used to generate the table is available by download (see Downloads).  Since there are in theory no practical differences between the revised Elo and the logistic Elo, the different results are likely due to a more precise definition of the derivative in [11.4].  Note that by taking the inverse of the derivative,

[11.6]              dP / dD  =  [P (1 - P) . ln10] / 400 ,

and substituting for P by the logistic formula [11.2], we get

[11.7]              dP / dD  =  [(10-D/400) . ln10] / [400 (1 + 10-D/400)2] ,

which is Elo's Verhulst formula (45) [E1, 8.43].

 
Rating Change Produced by a Win for Various Rating Differences
       
  Revised Elo Normal Elo Logistic Elo
  D > 0 D < 0 D > 0 D < 0 D > 0 D < 0
0 6.95 6.95 8 8 8 8
20 6.57 7.37 7.55 8.45 7.54 8.46
40 6.23 7.85 7.10 8.90 7.08 8.92
60 5.93 8.38 6.67 9.33 6.63 9.37
80 5.67 8.98 6.24 9.76 6.19 9.81
100 5.43 9.65 5.81 10.19 5.76 10.24
120 5.22 10.41 5.39 10.61 5.34 10.66
140 5.03 11.25 4.99 11.01 4.94 11.06
160 4.88 12.20 4.54 11.46 4.56 11.44
180 4.71 13.27 4.18 11.82 4.19 11.81
200 4.57 14.46 3.82 12.18 3.84 12.16
220 4.45 15.80 3.49 12.51 3.52 12.48
240 4.35 17.31 3.17 12.83 3.21 12.79
260 4.25 18.99 2.86 13.14 2.93 13.07
280 4.17 20.89 2.58 13.42 2.66 13.34
300 4.09 23.01 2.32 13.68 2.42 13.58
320 4.03 25.40 2.06 13.94 2.19 13.81
340 3.97 28.07 1.84 14.16 1.98 14.02
360 3.91 31.07 1.63 14.37 1.79 14.21
380 3.86 34.44 1.44 14.56 1.61 14.39
400 3.82 38.22 1.26 14.74 1.45 14.55
420 3.78 42.46 1.10 14.90 1.31 14.70
440 3.75 47.21 .94 15.06 1.18 14.82
460 3.72 52.55 .83 15.17 1.06 14.94
480 3.69 58.54 .72 15.28 .95 15.05
500 3.67 65.26 .61 15.39 .85 15.15
520 3.65 72.80 .53 15.47 .76 15.24
540 3.63 81.26 .45 15.55 .68 15.32
560 3.61 90.95 .38 15.62 .61 15.39
580 3.60        101.40 .32 15.68 .55 15.45
600 3.58 113.34 .27 15.73 .49 15.51
620 3.57 126.75 .22 15.78 .44 15.56
640 3.56 141.79 .19 15.81 .39 15.61
660 3.55 158.67 .16 15.84 .35 15.65
680 3.54 177.60 .13 15.87 .31 15.69
700 3.54 198.85 .11 15.89 .28 15.72
720 3.53 222.69 .08 15.92 .25 15.75
740 3.52 249.44 .06 15.94 .22 15.78
760 3.52 279.45 .06 15.94 .20 15.80
780 3.51 313.13 .05 15.95 .18 15.82
800 3.51 350.91 .03 15.97 .16 15.84