Beating the closing line in baseball: How important is it?

[jgilmartin]

The guy quoted an article arguing against his point of view as "conformation" for his point of view; do you really expect him to have thought about the subtler nuances of his position?

Good point, although this exact argument (with many of the same participants) took place over a year ago, and the steam argument was brought up then so Hu has had plenty of time to come up with a good counterargument.


I don't think Hu actually believes what he purports to. I think he is just trying to get successful modelers to post their models' output in order to get an edge without doing any work. Clearly this is unlikely to happen given that the modeler would, in effect, be giving away thousands of dollars.

[jgilmartin]

What destruction are you talking about? Show me connection first!

Steam chasing is a post openers phenomenon. and does not require any handicapping skills at all.
It is a simple form of trend following based on a hope that trend is (and will stay) your friend, as they say.

OK, why is betting steam profitable then? Why not just bet the other side at the new price?

[hutennis]

OK, why is betting steam profitable then? Why not just bet the other side at the new price?

B/c it is unprofitable to argue with unbroken trend without really good reasons.

Steam chasers are not skilled handicappers. They just trend followers. They can not have a good reason to bet another side.
Once trend is broken, steam chasing is gone. And it happens all the time. You just conveniently chose not pay attention to these cases.

And BTW,

I think he is just trying to get successful modelers to post their models' output in order to get an edge without doing any work.

This is silly. I thought you are a bit more of a thinking person than this.

There is no edge until there is statistical significance.
But as soon as there is statistical significance the experiment is over and there are no more "freebies"

So no advantage can be taken by definition.

[Carl-Haakon]

[jgilmartin]

B/c it is unprofitable to argue with unbroken trend without really good reasons.

Steam chasers are not skilled handicappers. They just trens followers. They can not have a good reason to bet another side.
Once trend is broken, steam chasing is gone. And it happens all the time. You just conveniently chose not pay attention to these cases.

Those originating the steam move are skilled. Do you disagree with this?

[hutennis]

Those originating the steam move are skilled. Do you disagree with this?

Those who more often than not originate proven to be profitable steam moves are skilled.
They are those who can gather and analyze relevant information better than most.
There are very few people like this. I would say 1 in 100 probably.

[jgilmartin]

Those who more often than not originate proven to be profitable steam moves are skilled.
They are those who can gather and analyze relevant information better than most.
There are very few people like this. I would say 1 in 100 probably.

Fair enough. Then we agree, some people can create a better line than the openers.

[hutennis]

Fair enough. Then we agree, some people can create a better line than the openers.

No! I did not say that. Don't put words in my mouth.

Everything we are talking about happens post opener.

There is no reason to attribute documented post opener success to unconfirmed pre opener assumptions.

The only way to demonstrate superior pre open handicapping skills is to post clean, pre Vagas openers
that subsequently will BTCL with statistical significance.

[jgilmartin]

If no one can create a better line than the openers, then why not just bet the other side post-steam move?

[hutennis]

If no one can create a better line than the openers, then why not just bet the other side post-steam move?

First, there is a difference between "no one can..." and " I don't believe anyone can until it is demonstrated that anyone can..."
I don't believe. The same way as I don't believe there is a god or lochness monster.

Second, I'm sorry, I don't see any logic in your statement/question.
Where is the connection?

[jgilmartin]

If the opener was the sharpest possible line (in other words, if no market participants could set a sharper line than the opener), one could profit by betting against the market on any line that had move significantly enough to clear the vig, could they not?

[hutennis]

If the opener was the sharpest possible line (in other words, if no market participants could set a sharper line than the opener), one could profit by betting against the market on any line that had move significantly enough to clear the vig, could they not?

No, they could not.

We cannot attribute significantly enough post opener move to simply uneducated attempt to argue with an opener.
Sometimes it is and sometimes it is not. Without complete information as to nature of this move (which we simply don't have)
we cannot make a judgment.

To say "I don't know the reason for this move therefore it is a bad move" is to simply make an argument from ignorance.
Action based on ignorance will not be profitable

[jgilmartin]

Wow.


You've agreed that betting steam is profitable. Most major line moves are not caused by significant player injuries/suspension/weather forecast changes/etc/etc; they are caused by an originator betting the line.


The only reason betting steam in these situations is profitable is because, as a whole, on average, the originators of steam moves are better at setting lines than the oddsmakers (and/or market members that have moved the line since open) are.


You have stated that your reasoning for not simply betting the other side after a steam move is because 'the trend is your friend.' The trend is only your friend because the originator of the trend is sharp. If the originator's bet was a coinflip, the implied probability of the pre-move line would be closer to the actual probability than the post-move line would be. If the 'trend' was being created by those setting inferior lines, it would be +EV to fade the trend, given the amount of cents you would gain after a steam move. You would essentially be getting all those cents for free.

[hutennis]

Wow........

Wow it is.

A lot bad logic and unwarranted assumptions in that post.
I'll come back to it later, though.
Gotta go play some tennis.

[mcuni]

The p value for this year alone for rate of line movement accordance is 1.795615e-07, under a binomial test for significance. For a sample size of 10,000 plays on my validation set it would be infinitesimal. But there I go with all that voodoo again.

A possible counter-argument to discuss.

I took a simple starting pitchers performance estimator and built a simple model with it. Just 1 parameter. Using filtered 2010-2012 data (2133 samples) showed that this model beats openers pretty well. It filters out 1600+ samples, and for 500 other samples it predicts the correct line movement >64% of times. In average using this model I can beat the opening line by 0.2 runs, or 8 cents.

I tried generating random numbers within the range of predictions (5.85 to 11 runs). Different random numbers beat the openers 50-55% of time, so I can state that probability of success is roughly 53%. That gives p value of 3,38136e-07 - which is probability of obtaining at least as extreme results as these (>64%) for ~500 observations. So, it's definitely non-random result.

Now finances. ROI for such bets is as low as -12%. Profit is negative for both correctly and wrong predicted line movements. What is that if not the evidence that beating openers is not connected with getting profit middle-term (or is 500 stakes a long-term already?)

Or more accurate it will be - not any consistent beating the opening line makes you a long-term winner. What do you think?

[Carl-Haakon]

Edit: simple misunderstanding

[matthew919]

You'd have to give me more details about what the pitching stat you're using is, as well as what kind of model. You can do this via PM if you'd like to protect the information, but I can assure you it doesn't matter, because there has to be an error in your methodology. There's no other explanation.

First, a single parameter could never give performance at that level, unless you're training it to identify line movement as the response variable, and then testing it on the exact same data set. Even then, that does not explain the negative ROI for that subset of games. 500 is pretty close to long term, although still possible to end up in the red with maybe a 10-15% chance (just a guestimate).

When you say "as low as" -12%, are you implying that there is more than one ROI being computed? ROI should be profit/risked, which should be a single number. It applies to how much you expect to make on each wager. Something tells me you're computing it as the amount you won/lost at the end of the year, compared to your starting bankroll (i.e.you started with 10,000 and ended with 8,800 so youre calling that a -12% ROI), which is incorrect.

EDIT: Just to clarify further, my reasoning in the above is mostly based on the line movement values in cents. A 64% BTCL rate is great, but the average cent movement is what catches my eye most. 8 cents is a spectacular number for MLB totals, and you will win long term with that, guaranteed. 64% with 5 cent average I would not put as much stock in. Still very likely to win, but not nearly the same ROI as I'd expect for 8c.

[mcuni]

You'd have to give me more details about what the pitching stat you're using is, as well as what kind of model. You can do this via PM if you'd like to protect the information, but I can assure you it doesn't matter, because there has to be an error in your methodology. There's no other explanation.

No problem, I took xFIP season-to-date built on adjusted pitching stats. That means 1B, 2B, 3B and HR allowed by starter for every game within a season were adjusted to weather and park, at which they were allowed. After that I calculated cumulative numbers, got xFIPs for both starters for every game where starters had enough batters faced season-to-date. To estimate a total I just summed both xFIPs. I don't remember if I adjusted the sum by park and weather for current game, if not, then I should do this next time. One can easily use SIERA or bbFIP instead of xFIP here as those are well correlated with each other.

So, I got an estimation for total (Te) with average of 8.12 runs for 2010-2012 filtered data. I chose confidence interval (CI) of +/-1 run. Now if Te-CI was more than adjusted opening total then I bet on Over. If Te+CI was less than adjusted opening total then I bet on Under. Otherwise it's a nobet situation.

Then in case of betting Over I took (adjusted closing total - adjusted opening total). In case of betting Under I took (adjusted opening total - adjusted closing total). For 307 of 477 bets the difference was positive. Sum of differences gives 96.925 runs. So average run per stake is 96.925/477 = 0.2032 runs, or given 0.25 runs = 10 cents it should be equivalent to 8 cents.

Average juice for those bets was 0.(09) (kinda -110/-110 line). It implies -9.1% ROI, or -43.36 units of profit for 477 stakes. I got -68.55 units with this model.

First, a single parameter could never give performance at that level, unless you're training it to identify line movement as the response variable, and then testing it on the exact same data set. Even then, that does not explain the negative ROI for that subset of games. 500 is pretty close to long term, although still possible to end up in the red with maybe a 10-15% chance (just a guestimate).

Yes, I agree that one variable is very unlikely to perform well in this case. Especially given that I don't even use attack estimator which is important for totals. I will double check all the formula as it was just a fast calculation. Still haven't noticed anything wrong there..
Actually there was no training even, just applying a single parameter to a subset.

When you say "as low as" -12%, are you implying that there is more than one ROI being computed? ROI should be profit/risked, which should be a single number. It applies to how much you expect to make on each wager. Something tells me you're computing it as the amount you won/lost at the end of the year, compared to your starting bankroll (i.e.you started with 10,000 and ended with 8,800 so youre calling that a -12% ROI), which is incorrect.

Excuse my english, I wanted to underline that ROI is unexpectedly strongly negative. When I say ROI I mean sum of profits for separate games (total profit) divided by amount of bets (given that one bet = one unit). I believe there is no mess.
Negative ROI was not that surprising when I checked correlations. For Te it was only 9.6%, and for adjusted total it's about 25%. So correlation seems to me a better indicator of long-term winning than BTCL.

EDIT: Just to clarify further, my reasoning in the above is mostly based on the line movement values in cents. A 64% BTCL rate is great, but the average cent movement is what catches my eye most. 8 cents is a spectacular number for MLB totals, and you will win long term with that, guaranteed. 64% with 5 cent average I would not put as much stock in. Still very likely to win, but not nearly the same ROI as I'd expect for 8c.

Yep, but so far I just don't see it winning. I will browse thoroughly the calculations and probably will find a mistake.
I made another fast calculation - added XR as attack estimator and couple of other variables and did training with previous seasons. Got 0.18 runs (7.2 cents), -8.4% ROI, 61.7% BTCL for CI of 1 run, and for CI of 0.6 runs it's 0.18 runs, -10.2% ROI, 63.45% BTCL out of 777 games. Correlation Te/Total is 18.5% for these 777 games but it's far from being enough as adj OT/CT hits 29-30% of correlation to actual totals for this subset.

UPD: I found that I didn't account for returns, that is when actual total was equal to OT, I counted it as -1 unit instead of 0. That makes ROI -7.5%, which is still quite poor.

[mcuni]

Actually I think I understand the issue.

The problem is that beating the closing line is easier than one can expect. Actual probability that randomly generated number between 7.6 and 8.6 will beat the closer is about 59%! Totals just seem to converge to center.

Try this. Calculate adjusted totals from OL and CL. Generate random numbers ranged [7.6...8.6]. Bet Under when Random+1 < OL and Over when Random-1 > OL. Notice which way CL goes. My calculations showed that line will go your way 58-60% of time. Average move will be >0.13 runs, or >5 cents.

What it means re binomial test for significance - probability of getting 64% BTCL is not that low, it's about 1%.
It doesn't explain why high BTCL results in negative ROI. But I think it is not by accident.

[hutennis]

Actually I think I understand the issue.

I don't think you do!

If your bogus playing with math leads you to a bogus conclusion that completely contradicts common sense, as a rational person using critical thinking you must start looking for a reason why your conclusion is bogus and not happily announce that you found some kind of sports betting equivalent of perpetuum mobile.

It is painfully obvious, that neither reason nor critical thinking is something you pay any attention to or value in any way! And you are not alone.

Look. Let me prove it to you that weight of an elephant is equal to the weight of mosquito.

Let x = weight of an elephant, and y = weight of a mosquito.

Call the sum of the two weights 2v ;
We have x + y = 2v

From this equation we obtain
x - 2v = -y and x = -y + 2v

Multiplying these together gives
x^2 - 2vx = y^2 - 2vy

Adding v^2 to each side:
x^2 - 2vx + v^2 = y^2 - 2vy + v^2

Factor: (x - v)^2 = (y - v)^2 ,
Take square root: x - v = y - v

Adding v to both sides gives x = y.

There we go! The weights are the same!

Happy sports betting!

[Carl-Haakon]

/htt

go back home, people

[mcuni]

If your bogus playing with math leads you to a bogus conclusion that completely contradicts common sense, as a rational person using critical thinking you must start looking for a reason why your conclusion is bogus and not happily announce that you found some kind of sports betting equivalent of perpetuum mobile.

Mom. This thread has already proved yourselves being a skilled bogus conclusions maker. Do not be jealous, I won't bother infringing on your imaginary mental supremacy.

If you read my post twice you will probably notice that there is nothing about profitness of such random betting. Smart person would rather repeat the experiment or at least ask about ROI instead of making conclusions about perpetuum mobile. And the answer is that ROI is negative, which is in line with what I said before: that BTCL 60% of time doesn't guarantee profit.

Adding v to both sides gives x = y.
There we go! The weights are the same!

Killer maths. Now I'm lost.

[mcuni]

You'd have to give me more details about what the pitching stat you're using is, as well as what kind of model. You can do this via PM if you'd like to protect the information, but I can assure you it doesn't matter, because there has to be an error in your methodology. There's no other explanation.

Matt, I guess the problem is how I calculate BTCL rate. There are 3 possible outcomes: (a) line goes your way, (b) line goes against you and (c) line remains the same. When I calculated BTCL rate I didn't account for (c), I just did a/(a+b) to check how often line goes your way rather than against you (excluding unchanged lines). But there can exist two more interpretations of BTCL. (a+c)/(a+b+c) - to check how often you do not lose to CL. a/(a+b+c) - to check how often you outmatch CL (so that being even doesn't count).

It seems to me that when you're talking about BTCL you mean the last formula (the strictest one). In this case 64% transforms to 61% in my example as there are 27 samples (~5%) with zero line movement.

So my conclusion out of all this is that BTCL of 60% doesn't guarantee a positive return though generally lets you be even with bookmaker (considering juice), >60% results in growth of positive ROI. And 55-56% for BTCL is possible to reach by simply betting on convergence of total to the mean. This doesn't result in any profit though.

[matthew919]

Some of these details are not making sense to me, but it sounds like you may be on the right track. When you mentioned a single predictor i figured you were going to mention somthing useless like whip.
So my advice: I would strongly urge you not to go into any more detail about how you adjust for park, or any of the exact formulas you use. You might even want to edit some of your previous posts while you still can- i cant say for sure if your method is likely to be profitable, and it is a small sample size to boot, but that information is worth protecting, at least as a foundation going forward. Keep up the work and you will have a profitable model soon i think. Good luck.

[hutennis]

If you read my post twice ...

I don't have to read any farther than this

The problem is that beating the closing line is easier than one can expect.

This is nonsense! Whatever leads you to this conclusion is bogus!

The same way as "weight of an elephant is equal to the weight of mosquito" is nonsense and whatever I used to come to this conclusion
is bogus! (btw, anyone who have not figured out what is wrong with elephant/mosquito "killer math" at the point of finishing reading it has no business "creating models" at all. Not even remotely sharp enough!)

If you don't understand this simple logic - good for books! They got it way to easy though.

[mcuni]

The same way as "weight of an elephant is equal to the weight of mosquito" is nonsense and whatever I used to come to this conclusion is bogus! (btw, anyone who have not figured out what is wrong with elephant/mosquito "killer math" at the point of finishing reading it has no business "creating models" at all. Not even remotely sharp enough!)

Not to insult you but I can hardly imagine someone unfamiliar with that trick unless he is under 13. And as for beating the CL, I don't want to impose my views, just try it and you will see. Take for the rule betting on OL unders for total 9 or more and betting on OL overs for totals 7 or less (adjusted the way Matt described). It will facilitate your pockets but will fulfill you with a sense of accomplishment as the line will move your way somewhat more frequently than just 50/50.

So my advice: I would strongly urge you not to go into any more detail about how you adjust for park, or any of the exact formulas you use.

Thanks for feedback. And I will follow the advice.

[hutennis]

Not to insult you but I can hardly imagine someone unfamiliar with that trick unless he is under 13.

You'd be surprised how many "modelers" are still searching internet for a solution.

And as for beating the CL, I don't want to impose my views, just try it and you will see. Take for the rule betting on OL unders for total 9 or more and betting on OL overs for totals 7 or less (adjusted the way Matt described). It will facilitate your pockets but will fulfill you with a sense of accomplishment as the line will move your way somewhat more frequently than just 50/50.

Will it inflate my pockets if I do the opposite?
If it will not and it will end up losing both ways, what do you think it would indicate?

[mcuni]

Will it inflate my pockets if I do the opposite?
If it will not and it will end up losing both ways, what do you think it would indicate?

It's losing both ways for sure, juice is acid.

[Underwraps]

Bump

[HeeeHAWWWW]

It's not baseball, but here's an example from a tennis model run since the end of 2011....

Opening line: 1027-860, 156 pushes. w/(w+l) 54.43%, 89 units profit.
Closing lines: 990-901, 175 pushes. w/(w+l) 52.35%, 57 units profit.


All the usual caveats should be made, most especially that opening lines have *much* lower limits. The second set of lines - a few hours after openers but with higher limits - yield somewhere in the region of 53.75% iirc. Certain subsets (surfaces, particular weather, etc etc) the market movements are less sharp than others.

Notes: 1/4 Kelly was used, but with a fixed bankroll (bankroll growth deliberately being made a separate issue), and with a 12% edge prediction cap. As an aside, using this resulted in RoI x1.14 higher than flat staking. That gap increases with more aggressive staking plans, but the above is the balance of profit vs variance I prefer.

[theclutch7]

take the dog late the fave easrly!????

[theclutch7]

you got it! it will hit 75 percent

[theclutch7]

ny state of mind!!

[chunk]

take the dog late the fave easrly!????

Generally speaking, no. However, you could be on to something here.

[LT Profits]

take the dog late the fave easrly!????

Maybe 5 years ago but not anymore, although there still may be a slight bias toward what you said.