Calculating Edge Based on Line Movement Question

[MarketMaker]

Can overall edge be estimated using the average line movement of a persons plays?

Using the data below what are the possible methods for using this to estimate win rate going forward? What would the confidence level be? Assume these movements are basketball totals.

Line Move Count Sum _Average
Against ____23 __49 __2.130434783
No Move ____3 ___0 __0
With ______59 178.5 __3.025423729

[dialup_king]

possibly you can get a rough estimate with half-point calculator especially with bigger sample size. A 3-point move on one game may not be worth as much as 3-point move in another game. I don't think points are always linear in other spots also

[Justin7]

Yes.

What is the average move?

[MarketMaker]

Average move is simply the sum of total points of each type of move divided by the count of those type of moves.

Example: Out of the 85 bets 23 times the line moved against for a total of 49 points, for an average move of 2.130434783.

[Justin7]

Yes, but what is the average move for all 85 bets? (And the median is useful also)

[MarketMaker]

Below are the moves from the actual bets. If I understand what you are asking regarding average move then it should be as easy as

(178.5 - 49)/85 = 1.523

If I treat the moves against as negatives and then find the median it is 1.5.

What does knowing the mean and median move averages tell me as far as overall edge going forward?

Moved Against (median 2)
3
2.5
1
1
2.5
0.5
4
2.5
2
2
0.5
3.5
5
0.5
0.5
1.5
1.5
0.5
5
1.5
2
3.5
2.5

Moved With (Median 3)
1.5
3.5
3
3
1
2
3.5
2
7.5
2
3
2
9
4.5
5
5
8.5
6
1.5
6.5
0.5
1.5
1.5
4
1
6.5
4.5
0.5
2.5
1
3.5
2.5
2.5
1.5
6.5
0.5
3
3
4.5
0.5
3
2.5
3
5
3.5
4
0.5
1.5
2
0.5
3
3
3.5
1.5
0.5
3.5
0.5
1.5
4.5

No Move
0
0
0

[Data]

Yes, but what is the average move for all 85 bets? (And the median is useful also)

Good God...

MarketMaker, read everything that Ganchrow posted here. If you did, you would come accross of this thread: http://forum.sbrforum.com/players-ta...r-success.html and other threads on this subject.

[MarketMaker]

Thanks mister Data I will take a look at those.

One other related variable I wanted to include. How would the calculation of overall edge be affected by the the person (persons) actually betting the plays as opposed to if the line moves occured independently without participation in the market by the handicapper.

Maybe after reading Ganchrows posts I will be able to answer these questions myself.

[MarketMaker]

I think sports has a better bell-curve than finances.

As I said, Z-score is a starting point. Another test I would use is the "line movement" test. How much did the line move from the time you recorded the bet to close? What is your median line move? If you're seeing a median line move of 2%, you probably are a very good player.

This is the quote from another thread that spurred me to start this thread. I just finished digesting the thread mister Data referenced above and don't think z score is nearly as indicative as looking at line moves in relation to a persons plays. I can't find anything that defines how to use line movement to estimate overall edge. If it does exist could someone please post a link?

[u21c3f6]

I try to respond only to threads that pertain to something that I actually do. I do not take line moves into account for any of my wagering selections and I certainly do not have the advanced math skills that some posters have but here is my opinion.

If someone is successful using line moves, I sincerely congratulate them but here is the problem that I have with what you are attempting. You are trying to use line moves as if there is some hard probability for the success of your individual wager based on the amount of line movement for that individual event. That would be true IMO if you were choosing all wagers and/or if your wagers were truly randomly selected, then the wagers should conform to the overall average line move "value". However, once you select a non-random subset of wagers, this subset may or may not conform to the overall average line move "value". Your line move "value" IMO should be calculated based on an adequate size sample of your subset of wagers.

When I try to figure out whether I have a +EV wager or set-up, I base it on my track record and selection ability which indicates to me the "value" of what I am actually doing. If someone told you they made a wager at -110 ATS, theoretically that wager is -EV but not IMO if the bettor has a track record of 55% winning selections.

Joe.

[Justin7]

When I try to figure out whether I have a +EV wager or set-up, I base it on my track record and selection ability which indicates to me the "value" of what I am actually doing. If someone told you they made a wager at -110 ATS, theoretically that wager is -EV but not IMO if the bettor has a track record of 55% winning selections.

Joe.

You will benefit from studying this alternative approach. If I see a handicapper hit 55% over 200 plays, but is flat versus line movement, I would conclude that he is lucky, not good.

[Dark Horse]

Add action points to that.

[Data]

You will benefit from studying this alternative approach.

This approach will provide meaningful results if and only if the bettor plays the board (as numerous times pointed out by u21c3f6) AND all his plays are 1 unit plays (as numerous times pointed out by Ganchrow). Otherwise it is pseudoscience.

[dialup_king]

I use Bayes Theorem to judge handicappers in addition to closing line and other factors. The prior distribution for those who don't beat closing lines has a lower probability of winners.

[Atomicdog]

I prefer to use Bet jamaica Lines and Vegas insider lists most vegas and offshore books. I track it from the open and wait until I see the late move 20 min. prior to start time.

[MarketMaker]

Would there be any validity to calculating a z score for line moves? I don't really like the solution as it doesn't provide any weight to how much the line moved, but it could be as simple as looking at the cumulative binomial distribution of the line moves.

Using the example in my OP if we assume that the line should move either way an equal amount randomly and taking into account instances where the line doesn't move than the probabilty for blindly predicting line moves would be 48.2353%. Since the cumulative binomial distribution of someone predicting an event with the probability of success of 48.2353% correctly 49 times or better out of 82 (I disregarded pushes, not sure if that is correct) is 2.3886% then can it be said this would be their Line Move Z Score? So basically a handicapper could say I have a mean line move of +1.523 with a Line Move Z Score of 2.3886% (1 in 41.8).

I would rather find a solution that incorporates how much the line moves but this is the best I have come up with so far. Feedback is welcome.

[Justin7]

This approach will provide meaningful results if and only if the bettor plays the board (as numerous times pointed out by u21c3f6) AND all his plays are 1 unit plays (as numerous times pointed out by Ganchrow). Otherwise it is pseudoscience.

I'll have to strongly disagree.

A standard modeling technique is to pigeon hole. What is the differential between your model's line, and the market line? group a: 0-2.5; b: 3-5.5; c: 6-8.5, ... If your model is good, you will see an increasing return (with fewer and fewer plays) as the differential increases.

If you play the board, you're screwed. Most of the games are close enough to right that you'll lose. If you weigh the whole board when evaluating your model, you'll introduce so much noise to your good plays, you'll render the model useless.

[Data]

A standard modeling technique is to pigeon hole. What is the differential between your model's line, and the market line? group a: 0-2.5; b: 3-5.5; c: 6-8.5, ... If your model is good, you will see an increasing return (with fewer and fewer plays) as the differential increases.

True, but, in the statement above, where is the line movement that you keep talking about?

If you play the board, you're screwed. Most of the games are close enough to right that you'll lose. If you weigh the whole board when evaluating your model, you'll introduce so much noise to your good plays, you'll render the model useless.

True again. But if you do NOT play the board then your selected plays are biased and the line movement tells you NOTHING about the subset you selected. The line movement increasing efficiency is true either for all plays (hence play the board) or for a RANDOM subset.

[MarketMaker]

I'll have to strongly disagree.

A standard modeling technique is to pigeon hole. What is the differential between your model's line, and the market line? group a: 0-2.5; b: 3-5.5; c: 6-8.5, ... If your model is good, you will see an increasing return (with fewer and fewer plays) as the differential increases.

If you play the board, you're screwed. Most of the games are close enough to right that you'll lose. If you weigh the whole board when evaluating your model, you'll introduce so much noise to your good plays, you'll render the model useless.

I think the same can be said for probability of success in winning as probability of success in predicting line moves when it comes to modeling. Success in predicting line moves should increase as the differential in the model's line and the market line increases.

[Justin7]

Data,

Assume there is a true value, p for all games. Designate it the true odds of the home team winning (although you could make it the odds of a team covering, or any other wagering proposition).

If sports betting is an efficient market, and your model is accurate, you will see two observations. 1. The greater the differential between your model's output and the market line, the greater the line movement you will see; and 2. The greater the same differential, the greater your return will be.

A "purer" way is to use no subsets, and run a regression of return and line movement versus the differential... but that is just mental masturbation.

Maybe I am not understanding you. You state "your selected plays are biased" if I don't bet the board? I want a winning bias (in testing and application). Would you explain more about the bias that worries you?

[Data]

Maybe I am not understanding you.

It seems it is not just me. u21c3f6 in this thread and Ganchrow in this thread said the same.

[Justin7]

It seems it is not just me. u21c3f6 in this thread and Ganchrow in this thread said the same.

I'll respectfully disagree with them, and rely on logic. All lines are not equally good/bad. If you assume this (as does the play all advocacy), you're assuming all lines are.

I understand the risk of selection bias. As an additional test, what is the likelihood that for each subset, an increasing differential shows an increase in average line move and average profitability? If they were random, and you had 4 subsets, it would be 1 / 2^8. Yet, almost every single successful model will pass this test.

[Data]

I'll ... rely on logic.

You have my attention.

All lines are not .... If you assume this ..., you're assuming all lines are.

Well, I did not take you too long to contradict yourself.

I understand the risk of selection bias.

Good.

As an additional test, what is the likelihood that for each subset, an increasing differential shows an increase in average line move and average profitability? If they were random, and you had 4 subsets, it would be 1 / 2^8.

I tried, but I could not completely get through this.

Yet, almost every single successful model will pass this test.

Justin7, you are going to have to give yourself an infraction. If you defined a "successful model" as a model that predicts line moves then, by definition, not "almost" but "every single successful model will pass this test". I will consider this as another Freudian slip and an evidence that you actually understand feel that a successful model is (defined as) something else.

[Justin7]

Justin7, you are going to have to give yourself an infraction. If you defined a "successful model" as a model that predicts line moves then, by definition, not "almost" but "every single successful model will pass this test". I will consider this as another Freudian slip and an evidence that you actually understand feel that a successful model is (defined as) something else.

There are lots of indicators that a model might be successful. The most important one is: does it make money when you apply it? A vast majority of models can be identified by the tests we are discussing, but not all. It's good to know if your model is crap or gold before betting it, and that is what these tests address. Some sports betting markets are simply so inefficient, that a line move test is not valid. These markets typically include ones with low limits that are in the first year or two of existence.

Successfully predicting line movement is a very strong indicator. If 90% of all good models successfully predict this, and 90% of line movements are predicted by winning models, you can draw your own conclusions about the methodologies. The numbers (at least that I have obvserved) aren't that high, but they are high enough to draw conclusions.

[Justin7]

Justin7 said All lines are not equally good/bad. If you assume this (as does the play all advocacy), you're assuming all lines are.

Well, I did not take you too long to contradict yourself.

I see where the confusion is. If you assume the opposite - that all lines are equally good or bad, I would agree with the conclusions of you and the others you suggested I read.

If all lines were equally good, a player's edge would probably not be sufficient to overcome the house's edge. Fortunately, there are good lines and bad. If you reject this assumption -- that some lines are better than others (and consequently some bets are better than others), I cannot see how you could develop a winning methodology, unless you believed that all betting lines are bad. I don't believe the market as a whole is this bad.

[Data]

You will benefit from studying this alternative approach. If I see a handicapper hit 55% over 200 plays, but is flat versus line movement, I would conclude that he is lucky, not good.

There are lots of indicators that a model might be successful. The most important one is: does it make money when you apply it?

I cannot take this anymore. Good night.

[MarketMaker]

I see where the confusion is. If you assume the opposite - that all lines are equally good or bad, I would agree with the conclusions of you and the others you suggested I read.

If all lines were equally good, a player's edge would probably not be sufficient to overcome the house's edge. Fortunately, there are good lines and bad. If you reject this assumption -- that some lines are better than others (and consequently some bets are better than others), I cannot see how you could develop a winning methodology, unless you believed that all betting lines are bad. I don't believe the market as a whole is this bad.

My assumption is that lines are normally distributed in relation to whatever the true line is.

[uva3021]

its reasonable to assume line movement moves with or against you with 50/50 probability

each potential scenario however induces a continuous variable, in theory anyway, and serves as fitting the function of :



where the interval [a,b]--line converted to implied probability--corresponds to the interval [0,1]

[Justin7]

its reasonable to assume line movement moves with or against you with 50/50 probability

Huh? Are you talking about random plays, or plays from a winning model?

[uva3021]

well in order to gauge your model you want to disclose the nature of the model's performance, so you assume the model is unproven, or random, then find where the model has its advantages versus making random plays

so if your model 55% of the time produces plays where the line moves 'with' you (close - your line > 0), then its a successful model, given an adequate enough sample size

[Justin7]

I misunderstood you. Got it.

55% is marginal, depending on the market.

[u21c3f6]

well in order to gauge your model you want to disclose the nature of the model's performance, so you assume the model is unproven, or random, then find where the model has its advantages versus making random plays

so if your model 55% of the time produces plays where the line moves 'with' you (close - your line > 0), then its a successful model, given an adequate enough sample size

This proves to me that your model is successful at selecting a non-random subset of events where the line moves for you more often than not. It does not prove to me that your model is profitable. It may be, but it may not be profitable. If you keep getting the "better" price on too many of the "wrong" sides, you will not be profitable even though the line moved with you more often than not.

The simple fact is (and the hard thing to master) is that to be profitable (on a flat bet basis), you must collect more on your winning wagers than you lose on your losing wagers. That is what a successful model should prove IMO regardless of whether it beats the "line" more often than not.

Joe.

[JustinBieber]

Why would 55% be marginal, isnt 55% really good for a model?

[Justin7]

Why would 55% be marginal, isnt 55% really good for a model?

55% win rate is great. 55% predictor rate of line movement is marginal.

[JustinBieber]

Isnt this basically the same thing in a market like NFL?