Question about "P(off)" in Justin7's "Conquering Risk" book

[brettd]

Out of Justin's "Conquering Risk" book (pg104):

"2. A team's offensive rating predicts how many points it will score (on average) against an average defense. This rating is measured by its average points scored per game in its first N games. You will call this P(off)."

Is this a fixed or rolling P(off) for the rest of the season? Say in round 5, we have an average of Team A's score for the first four rounds, which is 20. Is 20 then Team A's P(off) for the whole season?

Or do we calculate a rolling P(off) by averaging the four rounds previous to Round x?

[brettd]

If P(off) is fixed from round x until the end of the season, surely some consideration must be placed as to whether the rounds that made up the average for P(off) were played at home or away. In my dataset, I have some teams that played three of the first four rounds at home, whereas some teams are the opposite.

I know the league average HFA in this particular league is 6 points. Would I allocate a +3 to away team's score and -3 to home team's score for the first x rounds, and then calculate P(off)? Or would factoring HFA come in later?

Any advice/help would be much appreciated.

[Justin7]

After 4 weeks, that is your predictor for the 5th week.

After the 5th week, you have another week of data. When loooking at the 6th week, you use the first 5 weeks.

If you had to set lines for the entire season after just 4 weeks, you could use it that way... but lines made with more information will be better.

There are different ways to account for HFA. Your method is one valid way (although I don't know any football leagues with an HFA of 6).

By the way, the points-based model is the weakest one of all in that book. If you are doing football, I would focus on total yards or yards-per-play. Or, do a 3-model hybrid.

[brettd]

Thanks for the reply Justin. This is my first foray into modelling and so I'm keeping it simple to begin with. However a few questions just to clear things up:

After 4 weeks, that is your predictor for the 5th week.

After the 5th week, you have another week of data. When loooking at the 6th week, you use the first 5 weeks.

So just to be 100% sure, p(off) is an accumulated average game after game until the end of the season. So round 23 in my case will have averages from the last 22 games?

By the way, the points-based model is the weakest one of all in that book. If you are doing football, I would focus on total yards or yards-per-play. Or, do a 3-model hybrid.

I'm capping second division Australian football leagues (the WAFL and SANFL to be exact). Average points scored/allowed per game is roughly 98 points. The statistics are completely different to any other form of football, so unless you know about handballs and centre clearances, I'm on my own.

I really can't see why the points based model wouldn't work here though? The logic behind it is easy to follow, and I've already got a number of ideas of how I can explore honing the basic formula with my stats (such as kick, mark and handball differentials to name a few). This is a very niche market (and thus I hope, very inefficient), so I reckon I could be a chance in having the model being +EV.

Also would I calculate say, "P(kick)":- the kick differential; in the same way I would calculate P(off)?

Thanks heaps for your advice. Cheers.

[aqaz69]

is this the book from sbr store?

[Justin7]

Yes, average of all games.

I don't know enough about this league to comment intelligently, sorry.

The book is in the SBR store on amazon, and Google Ebooks.

[B1GER1C828]

Justin you needa make mroe appearences in SBR chat.