I thought I’d briefly revisit the results of the BeatPower confidence method of ranking picks. As you can see, the picks the model had higher confidence in all went as predicted. All of the incorrect picks were clustered in the bottom half of the chart, comprising all of the low confidence picks.
| Matchup (Winner-Loser) |
“Confidence” (out of 100) |
Result |
|---|---|---|
| Pittsburgh-Cincinnati | 83.9 | Correct |
| Tampa Bay-Detroit | 80.0 | Correct |
| Dallas-San Francisco | 71.8 | Correct |
| Washington-Seattle | 67.9 | Correct |
| Chicago-St. Louis | 67.3 | Correct |
| Minnesota-Jacksonville | 42.3 | Correct |
| Indianapolis-San Diego | 27.1 | Correct |
| New England-Miami | 15.0 | Correct |
| Buffalo-Kansas City | 13.9 | Correct |
| Arizona-NY Giants | 12.3 | Wrong |
| Carolina-Atlanta | 8.9 | Wrong |
| Houston-Cleveland | 8.1 | Correct |
| Denver-Oakland | 5.4 | Wrong |
| Green Bay-New Orleans | 5.3 | Wrong |
| Tennessee-NY Jets | 2.2 | Wrong |
| Baltimore-Philadelphia | -9.2 | Wrong |
It’s interesting to note that the standard Beatpath model correctly picked the Baltimore-Philadelphia, whereas the BeatPower “confidence” model got it wrong. Perhaps when they disagree, it may be better to go with the standard rather than the BeatPower method. Either way, the Eagles deserved to lose that game.
This method has grabbed my attention. It’d be interesting if someone could go back through the rest of the season and determine what these picks would have been at the time along with what the results look like.
I suppose you could go as far back as there are archives of the BeatPower rankings.
If this isn’t in the works, and nobody else gets around to it, I might toy with the idea. Just to see where this method really stands.
Great stuff as always. This site is fascinating.
Something’s a little odd about this. You’d expect the wrong answers to be clustered lower down rather than higher up, and so they are; but I fail to see why the low confidence picks should be almost entirely wrong. Low confidence means two teams evenly matched, which should mean even odds of right or wrong on any given guess, no? Why is picking at low confidence apparently sizably *worse* than coin flip? (Yes yes, sample size blah de blah blah blah, but if this pattern holds for more than a couple of weeks there’s something to look into there.)
Silent Speaker, I think what we’re seeing here is simply one week’s result. I’ve done this type of “confidence” ranking for a few previous weeks, and while the incorrect picks were still clustered toward the bottom of the chart, this week seems to be exceptional in this regard. Usually it’s more of a 2:1 ratio between incorrect picks in the bottom half versus the top half of the chart.
You can see my previous “confidence” results for Week 9 and Week 10. Below are the results from Week 11:
Carolina-Detroit: 100.0 (100.0-0.0) CORRECT
Tennessee-Jacksonville: 88.1 (100.0-11.9) CORRECT
Arizona-Seattle: 83.3 (90.4-7.1) CORRECT
Pittsburgh-San Diego: 78.1 (89.5-11.4) CORRECT
NY Jets-New England: 68.4 (100.0-31.6) CORRECT
Atlanta-Denver: 61.6 (89.6-28.0) WRONG
Philadelphia-Cincinnati: 59.1 (65.0-5.9) TIE
Miami-Oakland: 35.6 (40.4-4.8) CORRECT
Tampa Bay-Minnesota: 27.4 (79.6-52.2) CORRECT
NY Giants-Baltimore: 26.2 (80.0-53.8) CORRECT
Indianapolis-Houston: 19.6 (63.8-44.2) CORRECT
Washington-Dallas: 10.5 (92.6-82.1) WRONG
San Francisco-St. Louis: 6.5 (6.5-0.0) CORRECT
Kansas City-New Orleans: -9 (10.0-19.0) CORRECT
Green Bay-Chicago: -1 (70.4-71.4) WRONG
Cleveland-Buffalo: -9.2 (14.7-23.9) WRONG
As you can see, it’s not always as clear at Week 12′s results.
I tried looking into the close games to see if homefield advantage could be significant, but it doesn’t appear to have a decisive effect either.
I did some backtesting a couple of years ago and experimented with flipping the picks if the favored team was visiting. I found that if the home team was within 2-3 ranking slots of the higher-ranked visiting team, that flipping the pick might lead to a slightly better pick record overall. Gotta backtest further though.