Weather and Home Field Advantage
Someone recently pointed out a study that indicated home field advantage (HFA) is not the same for every stadium. While that's certainly true, it's very hard to quantify. By definition, the same team is always the home team when measuring a particular location's HFA, so in any given year there would be a lot of team strength captured in a variable accounting for the field's HFA.
The efficiency model I've used includes a factor for HFA, but it is the same regardless of climate. This is the beginning of an effort to quantify the effect of climate on HFA and to see how much of HFA is due to climate differences and how much is due to other factors such as crowd noise, referee psychology, or travel.
The table below lists each home team along with their average December weather. Click on the table headers to sort
| Team | Avg Dec T | Avg Dec Wind | Wind Chill |
| GB | 29 | 10.5 | 19.7 |
| BUF | 36 | 13.3 | 27.3 |
| CLE | 37 | 12.1 | 29.0 |
| CHI | 37 | 11.0 | 29.5 |
| NE | 42 | 12.2 | 35.3 |
| KC | 42 | 11.2 | 35.7 |
| PIT | 42 | 10.4 | 36.0 |
| NYG | 44 | 10.8 | 38.3 |
| NYJ | 44 | 10.8 | 38.3 |
| CIN | 44 | 10.2 | 38.5 |
| PHI | 44 | 10.1 | 38.6 |
| DEN | 44 | 8.4 | 39.3 |
| SEA | 46 | 9.5 | 41.3 |
| WAS | 46 | 7.8 | 42.0 |
| BAL | 49 | 9.3 | 45.0 |
| TEN | 49 | 8.9 | 45.2 |
| CAR | 54 | 7.4 | 51.9 |
| SF | 56 | 7.1 | 54.4 |
| DAL | 57 | 10.8 | 54.5 |
| OAK | 58 | 7.1 | 56.8 |
| HOU | 65 | 8.0 | 65.0 |
| ARI | 65 | 5.1 | 65.8 |
| JAX | 66 | 7.8 | 66.3 |
| SD | 66 | 5.6 | 66.8 |
| TB | 72 | 8.4 | 73.5 |
| MIA | 75 | 9.2 | 77.1 |
| ATL | 70 | 0.0 | 79.2 |
| DET | 70 | 0.0 | 79.2 |
| IND | 70 | 0.0 | 79.2 |
| MIN | 70 | 0.0 | 79.2 |
| NO | 70 | 0.0 | 79.2 |
| STL | 70 | 0.0 | 79.2 |
It's not a surprise that Green Bay is coldest by far, followed by places such as Buffalo, Cleveland, and Chicago. Green Bay would even qualify as a cold climate through November with an average 36 deg wind chill. But I was surprised by how much colder (and windier) a place like Kansas City is than Baltimore or Washington. I'm still considering how to classify each city. Domes are easy, but where is the line drawn between cold and moderate? Should there be cold, moderate, and "warm" classes? For now I'd put the line between cold and moderate at 40 deg wind chill, between DEN and SEA. I'd also define warm weather teams starting at 60 deg between OAK and HOU.
Continue reading this article here.
4 comments:
One example of what I'm suggesting:
If I removed the HFA for all games so far this year, the model's accuracy would improve by four games (or by 4.2%). Four games that were "incorrect" so far would become correct predictions. And none of the correct predictions would change to incorrect.
Then in December, adding a stronger HFA factor to games played by warm weather teams at cold climates might further improve accuracy.
You're absolutely right. And I think your finer-grained approach would be an improvement over my original study. The weather factor is an inefficiency in the Las Vegas spread and other predictive models. But I've also found it hard to incorporate into a system. I've found issues with sample size, but I'm wondering if your model isn't better suited for a weather variable since you're able to filter the effect of HFA down to one binary variable.
Also, I'm not sure you could reasonably prove each stadium has its own unique HFA value just because of sample size. I found splitting the data by the 4 months of the regular season gave weird numbers for that reason. Splitting it 32 ways would be even worse.
I am wondering whether just classifying cities as warm/moderate/cold will provide a true indication of what you are trying to find out. I suppose if you did it over enough seasons it would be right. But, given that each team would play on average 2 Dec home games per year, there could be alot of noise. For instance, one Sunday in Dec in GB could be 48 degrees with little or no wind. While classifying the road teams as warm/moderate/cold would be the best way to do it, I think you need to use the actual weather conditions for the game to overcome the obvious sample size issues. NFL.com has the weather conditions on games going back to the 2001 season (temp, weather, wind, winddir, wind speed, and humidity). I have been slowly inputting this data into my game database (only made it halfway through 2001 & have kept up with the current season so far). If anyone is interested in collaborating and splitting up the seasons to speed up the process, let me know. In addition to answering your question, getting all of this data in a database would allow us to answer alot of questions (e.g. how point totals and weather conditions interact, do ARI & TB have a "warm weather" advantage in the early season when the heat index is above a certain number (and if this does exist, whether it is more pronounced in the 2nd half as opposing teams run down), FG kicking and weather conditions, etc.
Derek-Thanks for the link. I was looking for your post on weather earlier and didn't find it. I agree narrowing down individual stadium HFA divides up the data too finely for reliable estimates. (By the way, the last 2 weeks wrecked my accuracy performance!)
Will-That's true about individual game temps. Using just average temp classifications, I'd need a really big data set to calculate precise HFA for cold weather. But on the other hand, a good average baseline is a good start for estimating a modest "late season weather" factor for a prediction model.
For example, there are 21 games in the past 5 years of dome teams playing at cold cities. The dome teams won only 14% of the games. That's pretty extreme, and much of that result is likely due to the relative strengths of the various dome and cold teams over that period. (And thanks for the tip about the actual game weather--it won't be too hard to look up those 21 games, or for other case-types such as warm at cold or dome at moderate.)
But a regression that accounts for team strength might provide a reliable estimate of the real HFA in that situation, especially when we compare the results with those from other combinations: cold teams at domes, or cold team visitors at cold cities, etc.
PS One of my long term hopes is to code a program that is a game book reader. It could simply parse up a batch of game book files and spit out a data file of game attributes and play-by play data.
Post a Comment