Steamsharp Chronicles.

[MattyKGB]

Great replies, thanks. This is the kind of intelligent debate I haven't been able to find at other forums.

Sorry for debating multiple points at the same time but that's kinda the road we're going down so let's do it.

no Win Expectation is not win probability, by definition probability is defined through a probability distribution which is normalized. win expectations are not normalized and you might notice that sometimes they add up to greater then 100%. This is by intention because we are only interested in the number of random performances [consider a statistical set aka boxscore a performance] that a home team for instance might beat an away team from a vast cohort of known performances.

There is nothing wrong with that. The issue for me lies in the way you are using your WE. You are comparing WE (not a probability) to Pinny's implied odds (which ARE a probability if you back out the vig). Another way to put it is that your WE is the win expectation against a random team, while Pinny's odds are a reflection of the win expectation against the specific team that they're actually playing against. It's an apples to oranges comparison. I think you note this in your response to PLP as a limitation of your method, but to me it seems like a pretty big deal - something that could take an otherwise good model and make it give bad results. But your results have been great so maybe I'm wrong about this.

Pinnacle often is forced to over juice teams like oakland (the 1.334 i was talking about) because the public over 12-14 hours is dropping 98% of the liquidity on the ML. in this case we hammer the underdog because of the "black swan" affect which looks a lot like PLP's superb RL/O bets on teams like MIA or NYM at 5-1.

I strongly disagree with both of these points.

If the heavy favourites are "juiced" because of public action, then A) there would be a clear pattern of favs becoming heavier throughout the day, and B) enough sharps would take advantage of this pattern to return the odds to equilibrium.

I am very familiar with N.N.Taleb and have read most of his stuff. I even spent big $$$ to store my baby's umbilical cord stem cells because that is the "antifragile" thing to do. There are NO black swans (or at least extremely few of them) in baseball. To use Taleb's terminology, baseball is a game that is rooted firmly in Mediocristan. Pretty much anything that can happen on a baseball field has happened at some point in the past. Black swans are events that are not accounted for in models because they have never happened before. A heavy underdog scoring 7 runs is NOT a black swan - it's happened hundreds of times before. I think you are confusing the concept of black swans with simple variance.

but very very rarely. baseball teams at best win 100 games on the season which to me says anything above 62.5% implied Wper to BE is suspect at best.

I disagree with this too. If the Dodgers played against Houston 162 times a year and if they got to send Kershaw to the mound 162 times against Houston's #5 starter 162 times, it's likely that the Dodgers would win far more than 100 games. So it's quite possible that a team could have >62.5% win probability in one particular game with one particular starter vs one particular opponent.

this in my opinion is false. (please take with a grain of salt) the problem is that people magically think pinnacle is the true win probability to break even when they post lines. I can show you a direct proof by contradiction that this is just false.

I don't know what others think but by being on the other side (seeing how books servers work), I know the books are just reacting to liquidity and trying to heap up bets which have negative win expectancy so that sharp money is rendered ineffective and they will either win a crapload of money when the fave loses or payout next to nothing when they win due to hedges and smart pricing. they know the true WP as well as anyone and the closing lines are almost always not that. (at least we have deluded ourselves into believing that and we will gladly open a bottle of 25yr McClellan's with you in Van if you drop by to debate it)

Just send me the scotch in the mail
There is nothing in my question that requires the assumption that Pinnacle's lines are efficient.
Suppose that you (based on your own models and analysis, not Pinnacle) determine that both teams have the same win expectancy. You have the choice to bet the visiting team at -104 or bet the home team at -104. You bet every game, so you have to pick one or the other. Which one do you choose and why?

Thanks again for the debate, looking forward to your responses

[steamsharp]

Matty, some would say intelligent debate is itself a win, regardless of the outcome.

There is nothing wrong with that. The issue for me lies in the way you are using your WE. You are comparing WE (not a probability) to Pinny's implied odds (which ARE a probability if you back out the vig). Another way to put it is that your WE is the win expectation against a random team, while Pinny's odds are a reflection of the win expectation against the specific team that they're actually playing against. It's an apples to oranges comparison. I think you note this in your response to PLP as a limitation of your method, but to me it seems like a pretty big deal - something that could take an otherwise good model and make it give bad results. But your results have been great so maybe I'm wrong about this.

I think here you have to be really careful, i am guilty of glossing to quickly over things. key to your argument above is the notion that baseball is a noisy random process. We believe this 100%. on any given day any team can beat any pitcher in any circumstance. you can call the WE the "signal" and the variation about pinny's BE the "noise"

in our world, all we say about pinny is that their closing odds imply: the win probability needed to break even over time. aka 100/odds = implied BE. I would not call this the true WP for say dodgers vs cubs. all it is is some price that implies how many times you have to win to make $$$.

the WE us not a win expectation against a random team, we would define it as the expectation of a statistical configuration (the stats aka ops, hit% BB% RISP% etc we use a lot of sabers) producing enough runs that it overcomes other known (previously played) statistical configurations with known runs scored. we know deep down that plenty of teams can score 19 hits in a game and somehow only put up 3 runs. we are very concerned with how on field play translates to runs and therefore wins.

so when we are comparing this WE, some expected set of statistics to a cohort of past games, all we care about is the delta between the WE and the implied BE (WP needed to win $$$ over many games) the reason we can do this and make a buck is we ASSUME that baseball is stochastic. on any day John Danks can throw a Gem and beat up Verlander. WTF! I call this kind of thing a black swan (or gray swan a hybrid) type event because guys like Kershaw exhibit vastly superior data that in no way points to a 4-2 loss. there is just no way the past data in the previous 15 games points to that. that would define baseball data as "type 2"

the theory is that if you get enough signal, aka (WE - BE)pit on pitching we say signal should overcome the noise and that heuristic works (say 20% edge) because we beat the closing lines over time. you just keep adding up those fat 3.48 to 1 wins and avoid enough outright terrible underdog bets that you end up making dosh. in basketball this effect is tripled as teams like ORL pay sweet 8 to 1 bets and somehow pound on ATL or GSW at home. god only knows why.

to wrap this up, there is nothing wrong with comparing a WE (assuming you have a GOOD method to map statistical sets to WE) to pinny implied WP to break even (BE) its just a heuristic and it works pretty great! the problem is saying "what is the WP from that" as PLP correctly surmised. we are currently working on the bayesian math to do this where the teams strength is factored as a conditional probability constraint.

in english, we have no way to say "this game has this much signal WE compared to the randomness inherent in baseball outcomes VAR(WP) and so we should hammer it" we just do it by feel right now and it works, there is a positive correlation to ROI.

[steamsharp]

I strongly disagree with both of these points.

If the heavy favourites are "juiced" because of public action, then A) there would be a clear pattern of favs becoming heavier throughout the day, and B) enough sharps would take advantage of this pattern to return the odds to equilibrium.

I am very familiar with N.N.Taleb and have read most of his stuff. I even spent big $$$ to store my baby's umbilical cord stem cells because that is the "antifragile" thing to do. There are NO black swans (or at least extremely few of them) in baseball. To use Taleb's terminology, baseball is a game that is rooted firmly in Mediocristan. Pretty much anything that can happen on a baseball field has happened at some point in the past. Black swans are events that are not accounted for in models because they have never happened before. A heavy underdog scoring 7 runs is NOT a black swan - it's happened hundreds of times before. I think you are confusing the concept of black swans with simple variance.

I would say no to this because baseball data when used to forecast the days event is a lot like the turkey problem. when you are using the recency data on a team with a starter to guess the expected outcome in the future, that data often is completely out of whack with the real result. the relationship between the recent data (say the last 15 games or so) to the foretasted result is massively chaotic. we just don't have enough relevant data to call baseball "Mediocristan" what happens is the relevant data creates a probability surface where the 7-0 MIA win is so unlikely you can call that "extremistan" for that probability space, highly highly unlikely. this surprised me when I first started researching sports and data.

I agree though, using Taleb's cool work is probably bad, but its such a good lingusitic tool to talk about probability its hard not to

[steamsharp]

I disagree with this too. If the Dodgers played against Houston 162 times a year and if they got to send Kershaw to the mound 162 times against Houston's #5 starter 162 times, it's likely that the Dodgers would win far more than 100 games. So it's quite possible that a team could have >62.5% win probability in one particular game with one particular starter vs one particular opponent.

proof by contradiction,

Justin Verlander.

any given thing can happen on any given data against any given starter. but ill give you kershaw, i saw him pitch in L.A and he destroyed my giants like they were pansies.

[ProlinePlayer]

your calculation is not correct for WP and here is why:

A) each WE is derived from a different heuristic which is based on different data (home and away + team specific). creating a joint probability is bad because its like saying you want to predict the chances of a dice turning up 6 twice in a row except you use a wood dice on throw 1 and a lead dice on throw 2 and maybe the lead dice has 2 sixes on it instead of 1.

B) the right way to deal with this is to use bayes rule, as the WP calculated from the WE for CHC vs a random opponent is surely depending on that opponent being above average or below. the dodgers are probably above average. we have been painfully stupid here just looking at the deltas between the WE's and saying that delta has some kind of predictive correlation to winning if its high enough (it does) we are working on how to do WP properly but its hard!

let me know if you think this is valid.

Sorry but I don't really understand Point A. I don't see why the methods used to create the WE values matters. In the end they represent the same thing - the probability of winning against a random opponent. And one should be able therefore to manipulate them into different values.

Re point B. It would seem to me that if the opponents are random then it would also be sound to assume that the average strength would be just that - average. I realise that may not be 100% accurate but it should be close enough for practical purposes.

By not converting the WE values it seems to me that a couple of strange things end up happening. For example because the Cubs have a WE of 45 they are just going to end up being an edge any time you can get a Pinnacle price of +120 or higher - at that price you'll end up with a implied probability which the WE beats.

It seems that what you have done, and I would assume not by intent, is ended up with an analysis and determination of an edge through a process which completely ignores who the team is actually playing against.
Using the Cubs, their WE is based soley on their past result. As I understand the fact that they are playing the Dodgers is not a factor that affects the WE.
But here's the thing. In the spreadsheet from the determination of the WE to the edge, at no point in between is their opponent considered. Maybe I've misunderstood but it really looks like the determination of the Cubs edge is based on an analysis of the Cubs only. The strength of their opponent seems to not be a factor.

PLP

[steamsharp]

Matty to wrap up, well have to drink that scotch to debate the Pinny mechanics. if the square liquidity overcomes the sharp liquidity by several factors, then normal market effects apply and the price will shift according to the weight of money. pinnacle should work like this because they will NOT let you drop 100k down on a side unless its obviously negative win expectancy.

how about that scotch?

[MattyKGB]

Matty to wrap up, well have to drink that scotch to debate the Pinny mechanics. if the square liquidity overcomes the sharp liquidity by several factors, then normal market effects apply and the price will shift according to the weight of money. pinnacle should work like this because they will NOT let you drop 100k down on a side unless its obviously negative win expectancy.

how about that scotch?

I still want that scotch (Glenlivet 21 is the best I can afford!) but my point has nothing to do with the Pinny mechanics. Any time you are betting against a book that charges any vig whatsoever, the possibility always exists that you will not be able to be either side of a game profitably, if your model ends up close enough to the odds you are being offered. I can prove this mathematically

[steamsharp]

oh i see mattyKGB.

yes in the case of the close games where the vig creates negative expectation on both sides here is our heuristic. surprisingly it creates a small edge across 2150 games of baseball this season

A) bet the home team always unless B)

B) if the pitcher is a new pitcher, bet the experienced pitcher side. if both our rookies No Bet.

[steamsharp]

PLP, your response it thoughtful and I need a little time to be better at explaining my thoughts.

one thing I can answer is yes absolutely the process is determined at random without respect to dodgers batting. this is because we assume as a condition of optimization that the dodgers are a completely random team coming into play against a pitcher who has shown an ability to throw up an inning by inning set of statistics over 90 innings (usually) we treat the dodgers performance as a random walk of 9 progressive statistical vectors driven by some foggy/fuzzy process of the starting and relief pitching they are playing against. the juxtaposition here is that the dodgers batting will get their turn in the preceding WE calculation when we use only offensive statistics they put up against other teams.4 (WE-BE)'s will be the variables we use in a WP calculation, not the WE's alone.

that's why i don't like the method you tried, we assume the dodgers are a performance chosen at random from the deck so to speak, and we dont want to add in the condition of considering the dodgers batting. if we did we would use bayes rule but its not straightforward how to construct that formula from the dodgers batting statistical drivers.

also you have to take batting into account as the top half and bottom half of the inning are mutually exclusive processes. I see you calcd off the data with only pitching data, probably good to add batting in there as its an equally potent force to consider driven by different data sets. (last 90 innings starter pitched != to last 90 inning batting batted)

[MattyKGB]

I don't know, the whole concept seems bizarre to me...artificially limit yourself to an extremely small sample size, then profit from the potential for rare events that are not in your sample. I'd say you were nuts if not for your good results. Still interested in learning more about your WE models. I might even be able to help you apply your WE's in a more logical way to generate win probabilities (you mentioned Bayes a couple times, it's something I use all the time).