Prepare for tXtFL Bowl 2.0! While we wait for the AFC/NFC division championships to choose their winners, check out last year's bowl, tXtFL Bowl 1.0, based on tXtFL 0.9.1.

This year's bowl will take advantage of the recently released tXtFL 0.9.2. You can also view more game testing of v.0.9.2.

Our goal is to test our football game model...and make some fantasy predictions while we're at it. Feel free to download a copy of tXtFL and start making predictions of your own.

Results.

Methods.

For each game, we recorded the points at the end of each quarter. At the end of the game, we record and highlighted stats of major players before reverting the spec files to original form. A summary includes each game and mean scores from all games. Along the way, we made notes about potential tXtFL bugs, and we intend to continue analyzing the stats to generate more realistic game models.

Analysis.

Plaxico's Platitude

Plaxico was more accurate than most predicted—our sim included. The New York Giant's receiver Plaxico Burress predicted a score of 23-17 with victory in the hands of the Giants, and even this score overestimated the final tally. One number was right, the 17, but this score belonged to the Giants. And the really crazy thing? The Giants were on top. The real scores came out to be roughly half of the tXtFL predicted scores, and the overall winner came from the underdogged, unexpected, unprecedented Giants.

The lowest predictions from the simulator never quite brushed with the depth of the real scores, but we did see a couple games with NYG scores at 26 points, within a field goal of Plaxico's prediction, and one game where NWE output a modest 22 points. But even these scores overestimated the real things by over a touchdown.

Did something go wrong?

We're inclined of course to say that the Giants and especially the Patriots went wrong, but there's no room in football or simulators for a failure to learn from inaccurate predictions. One of the difficulties of the current tXtFL drafting system is that intangible skill levels, such as hands, feet, smarts, have to be hand-edited into individual players. Teams also have coaching values that require hand-editing, and to keep the simulator as objective as possible, we opted to leave the default values for the simulation. We also don't have any parameters available for, say, over-confidence.

It would be extremely difficult in our opinion to create such values without an objective team of observers who could enter such values. The user (eg you) of course is free to edit any of these values and might be inclined to customize players particular importance, such as those from your favorite team or for your all-star batch of pros.

Conclusions.

tXtFL 0.9.2 probably came close to predicting the score predictions for Super Bowl XLII, but not quite so close to the real scores. The simulator remains subject to skewing by the numbers, without adequately taking into account the irregularities of real-life football. The hand-editable player skill values and team coaching points has emerged as a possible measure to fill in the gap between objective number-crunching and subjective weighting but requires a broad knowledge of individual players and their teams. Future methods to simplify this process or to incorporate a wider variety of objective parameters would help improve the simulator's accuracy.