Edge Percentage

**CHUBNUT** · 04-05-11, 03:40 PM

If we take for granted your way of assessing the line is reasonable then its no surprise your lines are almost identical to to real line. If however your line is out by lets say 3% then why would you think you have an edge when all logic will tell you that you are missing some critical info not taken into account in your model.

Thats why models are nonsense and you start looking at weather forefasts. Winning at betting takes intuition and experience which isnt found in the bottom of a mcdonalds family offer or a sports book.

**bztips** · 04-05-11, 07:05 PM

Yes, you're either missing something or you have an edge

It all depends on the confidence you have in your model AND how efficient you think the market really is. Many in this forum have a belief that the closing line in literally every game, especially in a major sport like MLB, is "efficient" -- at least to the extent of representing the "true" mean of the probability of a team winning (after accounting for vig). Others (like me) believe that sometimes the line SOMETIMES really is just plain wrong for a given contest.

Speaking only for myself, out of 2,430 contests in an MLB regular season, it wouldn't surprise me if the closing line were "significantly" off from the "true" mean in a few hundred cases. But that's just me -- for many others, the number might be more like 10 or 20 and which would occur only due to very unusual circumstances such as injuries, unknown ability level of a starting pitcher with no prior MLB experience, etc.

Obviously, you can test out your model by tracking its performance with different degrees of purported edge. If your model really is any good, then its performance and the degree of edge should be very highly correlated.

**jgilmartin** · 04-05-11, 07:25 PM

When you are doing your backtesting, do you record your perceived edge? I do, and then I go back and look at games where my 'edge' was absurdly high, and write down anything out of the ordinary. Does one of the starting pitchers have very few MLB starts? Does one of the bullpens have extremely good or extremely bad stats? Are there unusual weather conditions?, etc. etc. You will likely be able to find shortcomings with your model, as trends will develop on games with a very high perceived edge.

**dirk93** · 04-11-11, 10:49 AM

So I tweaked the model even more and all of my lines are within cents of the real line. My edge is now firmly between 2-5% on games I would consider betting. If I can consistently be this spot on for creating my own line, is 2-5% a big enough edge? If anyone wants to see my lines for tomorrow I will post them later to show you.

**That Foreign Guy** · 04-11-11, 11:14 AM

Depending how many games it produces an edge 2-5% is perfect IMO, it's close enough that you're not assuming the rest of the world are completely wrong, but different enough that you have a positive signal.

**bztips** · 04-11-11, 11:18 AM

My models typically are not accurate enough for me to be confident in just a 2-5% edge; maybe yours are much better. Can you compute a confidence interval for your estimates?

For example, if a line is -120, corresponding to a winning prob of 54.5%, then if your model is giving you a 3% edge, that translates to about -128 (56.2%). Does the 90% or 95% confidence interval around your prediction include the actual line of -120? If so, then you can't be real confident that your edge really isn't zero.

Again, speaking just for myself, my models are nowhere close to having tight enough confidence intervals to allow me to bet with small edges like 3%.

**WendysRox** · 04-11-11, 04:08 PM

This is a question that I need to research as well. I'm one of those crazy idiots that believes that lines are not efficient, but we better not go into that now (it's like discussing politics at Thanksgiving dinner around here). But, my point is: I would probably be OK with higher edges than some people would, but even I am not OK with a 94% edge in any major sport!

Here's a couple of examples from today:

Team / Predicted score / Win Probability / Line / Edge
Tor / 6.05 / 78% / 150 / 94%
Sea / 3.09 / 22% / -165 / -64%

Clev / 5.05 / 45% / 120 / -2%
LAA / 5.68 / 55% / -130 / -2%

My short-term solution to this is playing 1/10 Kelly. But, all I'm doing is reducing the size of the profit/loss. I guess it's pointless to use kelly at all when the data you're inputing doesn't seem legit. I've got a lot to learn....

GL with your research!

**WendysRox** · 04-11-11, 04:30 PM

Ahh... nevermind. I just looked over my formulas and I'm using a bastardized version of pyth. I just used something quick and dirty while putting together my spreadsheet last week... Sorry to clutter up another thread with my idiotic post.

Just out of curiosity, what do you guys use to determine win%, given predicted score?

**Flight** · 04-11-11, 11:07 PM

All variants of Pythag (http://www.crashburnalley.com/pwl.html)

The standard error in my models is way above that being discussed in this thread. Or should I say, my confidence interval is more like p > 0.2

Here's a funny comic for stat guys

Significant

http://xkcd.com/882/

**Flight** · 04-11-11, 11:09 PM

Sorry, double post, forum is jacked up right now

**dirk93** · 04-18-11, 09:35 AM

Anyone know where to find team runs scored projections? I know there are a few at the Yankees Replacement Level blog, but are there any other sources?

**ProphetofProfit** · 04-18-11, 02:46 PM

First, edge is odds x win probability - 1 right?

Ok, at what edge do alarm bells go off for most people here? I assume that if your models spit out bets of edge 20% you're not going to believe them. Or 15%, or even 10% might be too much in some sports. So if we have these illogical edges popping up in places, what is the best way of dealing with them? One is to cap the edge at a fixed amount if using kelly, so everything above this value is reduced, and everything below is left as is.

Another is some form of transform that I don't know how to deal with. Say, multiplying the edge by a number less than 1 to account for uncertainty. No idea how to quantify this at all.

**Pro Handicapper** · 04-20-11, 08:17 AM

One thing you need to look at is whether or not your win percentage increases as you edge increases on a market that has the same odds.

For example, if betting on a -110 market, does your win percentage increase as your edge increases on this market? If the answer is yes, you generally have a model that you can trust.

Back testing is very important to answer these sorts of questions and to develop confidence in your model.

Good luck.

**ScreaminPain** · 04-20-11, 12:54 PM

Originally posted by dirk93

Anyone know where to find team runs scored projections? I know there are a few at the Yankees Replacement Level blog, but are there any other sources?

google: Cairo and Zips

**dirk93** · 04-24-11, 08:06 PM

So I computed a 95% confidence interval for my edge and it hovers around 3%. Does this mean I should only bet games that I compute to have a 6-10% edge (-3% means I would still have a 3-7% edge worst case scenario). If my edge is too high in the first place, doesn't that most likely mean something is wrong?

**dirk93** · 05-04-11, 04:10 PM

Thanks for the help so far. I have another question. I have two systems, both using the pythag method. To get runs scored, I use either projected team runs scored for the season or I use each team's lineup for a given night and project each player's runs scored if they played every game of the season.

My question is that it seems like the flat projection for runs scored predicts lines closer to the actual line. However, shouldn't specific lineups for each game be more accurate? I'm in the process of gathering info to make a 95% confidence interval for my edge for each method. Am I possibly messing up the individual players runs scored per season?

**strixee** · 07-26-11, 06:40 AM

I'd like to calculate my edge based on my past wagers. The problem is I have placed them at various odds from like -150 to +1000. It seems to me that the simplest way is to normalize the
odds by multiplying them by a coefficient indirectly dependent on the odds. And the calculation would be k= 1/(decimal_odds-1), so e.g. for 2.00 k=1, for 3.00 k=0.5, for 1.60 k=1.66
If I'd be taking vig into consideration, then there should be some adjustment like k=0.95/(decimal_odds-1).
In the case of quarter handicaps I'd further multiply the odds by 0.5 when the spread/ total hits the closest whole number and when whole number hcps will be hit, then I'll ignore them.
So if in the end I'll get numbers like +540 (won), -460 (lost), my edge would be 8%, right?
(540-460)/(540+460)=0.08

**Justin7** · 07-26-11, 11:01 AM

Originally posted by dirk93

Thanks for the help so far. I have another question. I have two systems, both using the pythag method. To get runs scored, I use either projected team runs scored for the season or I use each team's lineup for a given night and project each player's runs scored if they played every game of the season.

My question is that it seems like the flat projection for runs scored predicts lines closer to the actual line. However, shouldn't specific lineups for each game be more accurate? I'm in the process of gathering info to make a 95% confidence interval for my edge for each method. Am I possibly messing up the individual players runs scored per season?

Yes, a line-up based version should be better (as long as you are not relying solely on YTD stats for players).

**bztips** · 07-26-11, 03:46 PM

Originally posted by Justin7

Yes, a line-up based version should be better (as long as you are not relying solely on YTD stats for players).

Justin, I assume you're suggesting that player stats should be weighted -- declining as you move further back toward the start of the year. Would you suggest going even further, and incorporating (small-weighted) player stats from the prior year as well??

**Justin7** · 07-26-11, 08:17 PM

Originally posted by bztips

Justin, I assume you're suggesting that player stats should be weighted -- declining as you move further back toward the start of the year. Would you suggest going even further, and incorporating (small-weighted) player stats from the prior year as well??

I do not know the best solution. I'm just warning against using the current season's results, and assuming those will carry forward. I had the single worst losing streak ever with this approach about 6 years ago. Then about a year ago, I found a Ph.D. thesis that analyzed that approach, and concluded that nearly any approach would work better than that.

Perhaps some combination of preseason projections (or the last 3 years), with a very small weight put on the current YTD results?

**uva3021** · 07-26-11, 09:03 PM

just use ZiPS data rather than YTD

**evo34** · 07-27-11, 12:25 AM

Originally posted by Justin7

I do not know the best solution. I'm just warning against using the current season's results, and assuming those will carry forward. I had the single worst losing streak ever with this approach about 6 years ago. Then about a year ago, I found a Ph.D. thesis that analyzed that approach, and concluded that nearly any approach would work better than that. Perhaps some combination of preseason projections (or the last 3 years), with a very small weight put on the current YTD results?

do you have a link to that paper?

**Justin7** · 07-27-11, 12:39 AM

Originally posted by evo34

do you have a link to that paper?

No longer, sorry.

**Scorpion** · 07-27-11, 02:37 PM

Edge question, is 14 cents edge in money lines in MLB = 1 pt edge in spreads in NBA or CBB?

**Pancho sanza** · 07-27-11, 03:13 PM

Originally posted by Scorpion

Edge question, is 14 cents edge in money lines in MLB = 1 pt edge in spreads in NBA or CBB?

Can't look at it that way since 14 cents edge when the market line is -300 is not the same as when the line is -125 right?

Use win probability if you want to compare them.

**Tackleberry** · 07-28-11, 03:11 PM

Originally posted by Justin7

I do not know the best solution. I'm just warning against using the current season's results, and assuming those will carry forward. I had the single worst losing streak ever with this approach about 6 years ago. Then about a year ago, I found a Ph.D. thesis that analyzed that approach, and concluded that nearly any approach would work better than that. Perhaps some combination of preseason projections (or the last 3 years), with a very small weight put on the current YTD results?

What data would you use for rookies then?