Weather biases in the NFL totals market

Abstract

I examine outcome predictability in the National Football League totals betting market using data from the 1984 through 2004 seasons. Results suggest that while weather is an important determinant of scoring, the market does not accurately incorporate the effects of adverse conditions into totals bet prices. Specifically, I demonstrate that heat, wind and rain reduce point production, and provide evidence that bettors underestimate this effect. I also present a betting strategy that accounts for expected weather conditions and produces an out-of-sample win rate significantly above the 52.38% profitability threshold.

Notes

¹ For instance, from 1976 to 2002 the Tampa Bay Buccaneers, a team located in a hot and humid region, lost 21 straight games in which the kick-off temperature was below 40°F. From 1992 to 2002 the Green Bay Packers were quarterbacked by Brett Favre, who won 35 consecutive games when kick-off temperature was below 34°F.

² Using data from Tradesports.com, an online exchange listing futures contracts on NFL and other sporting events, I find that during the 2002, 2003, and 2004 NFL seasons, the ratio of dollars wagered on sides bets to that wagered on totals bets is 9.3 to 1.

³ While this has been the traditionally accepted mechanism, Levitt (2004) suggests the possibility that bookmakers have an incentive to leave the book unbalanced in order to maximize their profits. This and other similar findings do not detract from the main point that the line may move slowly enough that profits can be captured because weather conditions are rapidly revealed with increasing certainty as kic-koff times approach.

⁴ Sides and totals lines are typically posted five to six days in advance of each event.

⁵ For games played in domed stadiums, I set temperature, wind, rain, snow, and humidity to 72°F, 0, 0, 0, and 50% respectively. I also repeat each analysis dropping from the sample all games played in domed stadiums. Results remain intact.

⁶ Panels C and D include only those games in which measurable amounts of rain or snow are present.

⁷ Test results indicate that multi-collinearity is not a problem in this specification.

⁸ Here a probit specification is preferable to an OLS specification because the number of points by which a bet covers is relatively unimportant compared to whether it covers.