Quote:
Originally Posted by JSkelts
Also, has anyone worked with the models mentioned at the end of the article? Specifically the Hunter one?
I'm not familiar with Hunter's work with Bradley-Terry modeling specifically, but when I ran the snooker data through a Bradley-Terry model a week or so ago it did backtest with a positive ROI, just not as much as Elo.
Speaking of Elo, clearly it has performed pretty bad since yesterday. While asking myself if the ship has started to sink, I rank backtests with different lookback windows (excluding today's matchups) and had the following results.
recent 25 matchups:
n=24, accuracy=56% (market=56%), roi=-2.48%
recent 50 matchups:
n=45, accuracy=64% (market=54%), roi=+18.19%
recent 100 matchups:
n=82, accuracy=62% (market=49%), roi=+21.53%
recent 1000 matchups:
n=671, accuracy=65.4% (market=64.4%), roi=+10.72%
recent 10,000 matchups:
n=7,336, accuracy=66.88 (market=68.21), roi=+3.8%
In the recent 25 matchups, it isn't doing any worse than the closing lines at predicting winners, but isn't beating the juice either. J. Trump's recent performance alone has cost us several units. My take from the results above is that if we increase the sample size, the ROI should re-converge to a positive number, so I will continue using the current implementation of the model for now.