Note that 100 half wins is the equivalent of 50 wins, and 100 quarter wins is the equivalent of 25 wins. My rating for 100 full wins is 1722, for 100 half wins is 1733, and for 100 quarter wins is 1728. These are all above my actual rating of 1715, which is to be expected. But since I expected the rating for 100 wins to be highest, I rechecked my code and the results, but nothing was amiss. Greg's rating for 100 full losses is 1437, for 100 half losses is 1452, and for 100 quarter losses is 1459. This does follow the expected pattern of being lower for greater losses. Also, they are all lower than his actual rating of 1476, which is also to be expected.
Let's now compare these ratings to those from the 100 draws experiment. My rating there was 1675, and Greg's was 1494. If we take the 100 wins experiment as describing an unmodified set of scores, and we take the others as attempts to reduce the weight of old scores, this seems to be the least fair of all. But let me try another experiment. In this experiment, I will change 100 wins for me into 75 for me and 25 for Greg. In this one, my rating is 1699, and Greg's is 1465. In this experiment, my rating dropped further than it did in the one-sided experiments, and Greg's rose more than it did in the one-sided ones.
It seems fairer to just reduce points than it does to adjust both sides. This reduces the significance of wins without giving a loser any false wins. But by what factor to reduce the wins of a certain age is still a matter of arbitrary decision. I know of no objective reason to favor one way over another. Furthermore, reducing older wins has the same kind of effect as not counting wins older than a certain age. It just does it in a more complicated and less transparent way than filtering out games below a certain date. So, I see no reason to add the ability to reduce the values of older wins.
As a further experiment, I created three more ratings scripts with additional data:
100 additional wins by fergus against mageofmaple
100 half wins by fergus against mageofmaple
100 quarter wins by fergus against mageofmaple
Note that 100 half wins is the equivalent of 50 wins, and 100 quarter wins is the equivalent of 25 wins. My rating for 100 full wins is 1722, for 100 half wins is 1733, and for 100 quarter wins is 1728. These are all above my actual rating of 1715, which is to be expected. But since I expected the rating for 100 wins to be highest, I rechecked my code and the results, but nothing was amiss. Greg's rating for 100 full losses is 1437, for 100 half losses is 1452, and for 100 quarter losses is 1459. This does follow the expected pattern of being lower for greater losses. Also, they are all lower than his actual rating of 1476, which is also to be expected.
Let's now compare these ratings to those from the 100 draws experiment. My rating there was 1675, and Greg's was 1494. If we take the 100 wins experiment as describing an unmodified set of scores, and we take the others as attempts to reduce the weight of old scores, this seems to be the least fair of all. But let me try another experiment. In this experiment, I will change 100 wins for me into 75 for me and 25 for Greg. In this one, my rating is 1699, and Greg's is 1465. In this experiment, my rating dropped further than it did in the one-sided experiments, and Greg's rose more than it did in the one-sided ones.
It seems fairer to just reduce points than it does to adjust both sides. This reduces the significance of wins without giving a loser any false wins. But by what factor to reduce the wins of a certain age is still a matter of arbitrary decision. I know of no objective reason to favor one way over another. Furthermore, reducing older wins has the same kind of effect as not counting wins older than a certain age. It just does it in a more complicated and less transparent way than filtering out games below a certain date. So, I see no reason to add the ability to reduce the values of older wins.