alsi FWIW the idea in the Football paper is how I've made bank for run scored in first and HRE and is essentially the idea behind derivatives.

Essentially calculating P(there's a run scored in the first | full game total is X)

You may wonder why not use 5i total since that's closer to FG? Hard to get 50/50 due to the juice and the clustering of totals. Bigger spread of totals for FG (usually 6.5-10.5 minus coors) vs. 5i is usually 3.5-4.5

only difference is I adjust it for how old the data is (i.e., 5 year old data would have a different weight than 1 year old) and MLB has enough data we don't need to make assumptions about a distribution.

I think I have solved the eSports dilemma by tracking the last 72 games offered at 5D/BetOnline. Spoilered because that will protect us from the vultures.

In order to obsess over eSports, I set this up really quickly.

I will move it over to crowdsourcesyndicate.info soon and make it look nice but

historical eNBA odds for current [today and ones for tomorrow that are posted] games

http://172.104.214.127/eNBA.php

The fade the steam is 6-0 so far today. Looking like 7-0 with Detroit +1.5 (up by 11 in the 4th).

Based on the system, the next thing we should fade is the LAC steam and take WAS -9.5

I'll start keeping the results too once we get some data.

This is also the method I use when I get really annoyed with small cleanup items in python -- glad to hear I'm not too far off base!

Any sites posting lines on o/u racial slurs for the next eNascar race? Feels like a good spot to take u0.5

gotta be rough to ruin your career over exposing yourself as a racist. gjge larson

ok so i'm taking this eSports NBA to the next level.

I got automatic scraping of the scores, opening and closing Spread, Total, ML, and Team Totals.

The output looks like this

https://pastebin.com/nUrjzsyR

First couple of entries are missing TT and ML as I was setting it up.

I also got a database to track bets (it will appear at the bottom) as well as auto grading every morning. All put together in about an hour.

How (why) do we do it.

Now to do the same for NHL

ATL @ LAL (2020-04-14 @ 00:35) [ID: 624023473]
Totals (Total, Over Price, Under Price):
00:22 214.5 -115 / -115
00:19 212.5 -115 / -115
19:54 211.5 -115 / -115

guess we going under 214.5

not a bad start to the site

Total Bets:
2020-04-13 HOU @ BKN Risk $575 on Over 204.5 Result: W $500 Close Total: 204.5
2020-04-13 UTH @ ORL Risk $575 on Under 181.5 Result: W $500 Close Total: 181.5
2020-04-13 LAC @ WAS Risk $575 on Over 205.5 Result: W $500 Close Total: 206.5
2020-04-13 LAC @ WAS Risk $575 on Over 207.5 Result: W $500 Close Total: 206.5
2020-04-13 LAL @ GSW Risk $575 on Under 225.5 Result: W $500 Close Total: 224.5
Total Net: $2500 on $2875 risked (87% ROI)

Spread Bets:
2020-04-13 LAC @ WAS Risk $575 on WAS -9.5 Result: W $500
2020-04-13 DEN @ DET Risk $575 on DET 1.5 Result: W $500
Spread Net: $1000 on $1150 risked (87% ROI)

Do 5Dimes/BoL have APIs or are you scraping these?

Finishing up reading the logistic regression notes. Very nice explanations. Was confused about page 13 w.r.t the Bayesian explanation for L2. I'll have to read up on how Bayesian statistics works. But can I gloss over that part for now, or is it critical? I'm sort of content just thinking that adding a scaled version of ||w||^2 to the error will keep ||w|| small without going into all that Bayesian stuff.

Also, you mentioned smoothing parameter lambda depends on the data. Although there's no universal formula, are there intelligent ways you've found of guessing what that value should be? Would you basically try training with different lambdas and see what maximizes log likelihood on new data that the model hasn't seen? Kind of a manual gradient descent.

Yep the idea behind Bayesian (and using training data in general) is that you have a prior which is I believe the data behaves this way based on what I know about it. So L2 sort of assumes that prior is normally distributed. The notes about why we divide (which should be multiply) by 1/sqrt(d) make a case for why we should add normally distributed noise to the data to minimize the cost, so that's basically it.

Then sort of as we collect data we update our prior beliefs. Bad prior belief? Bad data. Good prior beliefs? Data will make sense faster.

Yep usually you just try different numbers out. Some packages in python have algorithms they use to determine the "optimal" number but sometimes they go crazy and produce non sense answers. The issue with gradient descent is the step size. Too small might never find the dip or find a local min instead of global. Too big might step over it. There are better techniques that address this, but they're not perfect in every scenario and work on every data.

It's kind of like k-means clustering. How do you choose a k? Just make a hockey stick (or elbow) plot. How do you do that? Solve the cost for a bunch of k's and plot k vs. cost. Yea but what if each k takes days to calculate? Then use k-folds cross validation and hope it works for your data.

5D has a XML file if you're an affiliate. Just make a site, never refer anything to them, and use it forever. Betonline used to make their odds available on their site as a XML file but it's gone.

Hit a rough patch with the eSports last night losing 3 in a row (!), but we came back and won 5 in a row before losing again. I made some minor changes because that TT thing was not working for some reason, but it should be fixed. Right now looking like 8-3 on totals and 4-1 on spreads since we set the site up.

I made one for NFL too

http://172.104.214.127/eNFL.php

The data coming for NHL is kind of messed up because the book is offering +0.5 at so many different prices (which is an error, they have things like +0.5 -110 and +0.5 -250 for same game). So I'll skip that for now, mostly because NHL is stupid.

Once we got a couple weeks worth of data I'll move it over with historical odds and such. Make it look nice.

FYI,
Zack Baun tested positive for a diluted sample at the Combine. You can still get him not to go 1st round at -140.

The testing positive for weed or whatever isn't the issue here; it's that Baun needed to add water weight just to hit 238 lbs. Has to be concerning for teams trying to project who are already torn whether he is an edge rusher or off-ball LB.

sigs, i'm surprised you don't put your chops into the markets instead of what is likely comprosmised sim sports

which is more compromised - russian eSimulations or the stock market?

i just wrote a big quarterly taxes check and sent it off to the govt to give to boeing to inflate their stock price

this is more to exploit what i think is a flaw in certain books vs. actually beating eSports sims.

past 24 hours eSims
+$5550 on $13800 risked
nearly 40% ROI

should we start eSims touting? how are we not cut off yet?

Is anyone betting these eSims directly on 5d? I can't bet there, but I'm curious why some times they move 3-4 points and other times only 1. They can't really move 4 points just on a max bet can they?

Football went a great 3-0-1 today.

Running at about 40% ROI over 20k bet, is this sustainable? How are we doing this? Did we solve it or are we running insanely hot?

Losing by 1/2 a point is rough in eSims tho. One was a meaningless basket by a team down 12 with 4 seconds left, Other was a team who was down, had the last possession and decided to dribble it out instead of trying to score, third was team who was up got a rebound with 20 seconds and the other team didnt' foul and just let them dribble it out even tho it was close

feel like this phoenix +12.5 is setting up to lose by 13.

The L1/L2 discussion sort of opened my eyes to something. I can modify the cost function whatever way I want (as long as I can still take the derivative). So if I want to weight more recent data more heavily, I could e.g. scale down the error based on how long ago the thing happened.

eSims tout catchphrase?

i'm worried.

went to sleep and since i woke up there's been 0 line moves in eSims NBA

have they figured out the scam? the last line NBA move was 5:03 eastern, aka 8 hours ago

18.5 point underdogs Pelicans up by 4 over Bucks with 7 minutes left. Fading the Bucks has been super profitable recently.

Found a eSports Sims Tout, we can steal their gimmick

https://twitter.com/athleteentry

lol this e football is hilarious

https://clips.twitch.tv/EnjoyableSpookyDogJebaited

of course we have the over

book i use just started offering ebaseball, lets get it

With SAC covering the spread and hitting the under we are now +11.5k off $500 bets in 48 hours.

I'm 99% sure this is rigged now.

fading the steam 5-0-1 in NFL

15-6 NBA totals
15-4 nba spreads

35-10-1 lol

only explanation is they hope people will chase the steam so they artificially create it.

Sigs before this gravy train derails- I think the php script has Phi +11.5 in the next game instead of -11.5?

got any NBA picks for the Goldensim?

Sigs shot you a pm but something you should probably take note of: 5dimes is offering two different NFL sims. twitch.tv/sportsbooktv are likely the ones you want to keep track of. twitch.tv/primetimeaction are being offered by CRIS with higher limits. I would venture to doubt that CRIS is doing any shady in comparison to 5dimes. May be worth distinguishing results between the two streams.

This was my favorite sim "highlight" https://www.twitch.tv/livefootballsi...akingLorisMau5

haha dak is amazing

oh yep phi fixed it when i added on por +13.5 too, so gonna get double loss there. por is keeping it sorta close, only down by 8. rough loss on MINN tho with a score as time expires to lose by 1/2 point.

lost first NFL game too, sad days.

the two football is interesting because they seem to have similar totals and spreads but one is 10 minute quarters and the other 15.

the one that was 15 had something shady where the score was 0-0 at half, they "lost connection" for 5 minutes during half time and then when they came back the teams went crazy and scored 70 points in the 2H. Seems legit