Resetting every season

I think resetting the model every season was perhaps a mistake. It makes it too sensitive to periods of unusual results. Now that I have data from several seasons, it's more feasible to experiment with using the last few seasons for prediction.

But I still think using too many past seasons should be avoided. My feeling is that the game changes pretty quickly, so the factors which will predict results will also change pretty quickly.

Past seasons

After tweaking my model for the 2009-2010 and 2010-2011 seasons, I tried it out on previous seasons. It showed mixed results:

2003-2004 reasonable profits
2004-2005 good profits
2005-2006 big losses
2006-2007 reasonable profits
2007-2008 small losses
2008-2009 good profits
2009-2010 huge profits
2010-2011 big profits

I need to decide how much these results matter to me. Of course I would have been happier if every season had showed a profit, but the 2005-2006 season was five years ago, which is a long time.

What data is valid?

How should I decide what data to include in my analysis?

Is there most similarity between games in a particular season? A particular league? A particular team?

Is there more similarity between a 2009-2010 Premier League match and a 2010-2011 Premier League match, or between a 2010-2011 Ligue 1 match and a 2010-2011 Premier League match, or between a 2010-2011 Premier League Chelsea match and a 2010-2011 FA Cup Chelsea match?

So far I have been analyzing data by season. I reset my analysis at the beginning of each season. But within each season, all teams contribute to the model. Is it really reasonable for a Blackburn-Bolton match to affect the model I use to predict a Spurs-Fulham match in the same season?

Positive profit

I've found a method that shows profit for both the 2009-2010 and the 2010-2011 seasons. Unfortunately the variance is still pretty high - at one point in the 2010-2011 season, the account shows a large net loss, before recovering to show profit. So now I want to reduce the variance.

It's hard to know whether two seasons of data is sufficient. On one hand, it's not a lot of games, making it hard to separate real trends from random noise. On the other hand, it seems likely that the sport of football changes quickly, and that older seasons are less similar to the upcoming season than recent seasons are. So perhaps it's not wise to analyze too many past seasons.

Predicting football

I'm a software engineer with interests in statistics and betting. I've been working on a personal project to predict English Premier League football results using AI. I've been working on data for the last two seasons and developing algorithms to decide when and how to bet. If I can develop the system to the point where it shows a profit for the last two seasons, I'll run it live on the upcoming 2011-2012 season, posting the predictions and results here.