that provider doesn't actually originate lines themselves FWIW, they require a minimum amount of books who they follow that offer it and then take a consensus line from those, with some settings we can adjust for how many/what kind of books, what kind of consensus, how much margin, the limits, etc.

Good to know. I have interviewed (unsuccessfully lol) for a few trader jobs in the past year, and got the impression from the technical assessments I was given as well as from conversations with senior traders that the US based books rely more on a linesmaking provider for certain markets or will price up more markets in house depending on the size of the business.

Is that observation consistent with your experience actually working as a trader?

well we aren't a US book so I don't fully know how they operate but given most of them have Euro providers that's a pretty fair assessment, whereas we are a Euro book so we do a lot of the stuff in house ourselves that they are getting from elsewhere but then because of the sheer size a lot of them are, they are able to price up a bunch more side markets themselves.

that's why u see some of the US books with 2 sets of props, one coming from a bet builder and the other in-house or from another provider with many of the same players and similar markets offered in each but betbuilder also provides the ability to parlay a lot of these player markets

While we're waiting for the results of the Si v Hendry matchup, here's what I have for tomorrow.

pengfei v castle -+357, castle +425
sharav v hicks +-121, hicks +149
carrington v obrien -+346, carrington -225
sijun v lawler -+192, lawler +329

No love from Hendry today.

Today:
0-1 -1.00u on 1.00u risked
Total:
17-16 -3.93 on 46.48u risked

today:
2-2 +2.29u on 5.25u risked
total:
19-18 -1.64u on 51.73u risked

Snooker plays for Friday's matchups:
williams v figueiredo -+185, figueiredo +207
hamilton v mcleod -+383, hamilton -274
honghao v lines -+162, lines +235
greene v cahill +-177, cahill -166
grace v hallworth -+119, hallworth +130
wells v jones -+195, no bet
ursenbach v kleckers -+262, kleckers +408
page v filipiak -+237, no bet
craigie v hugill -+545, no bet

Snooker plays looking good so far this morning. I'll finish grading the results tonight but in the meantime here are the numbers for tomorrow. The model is pretty close to market numbers on most so I only found 3 plays.

burns v lichtenberg -+172, lichtenberg +188
zifan v mann -+115, no bet
zhou v si -+403, no bet
donaldson v dale -+191. no bet
slessor v highfield -+146, slessor -122
selt v bond -+365, selt -298
milkins v oconnor -+147, no bet
maflin v jones -+162, no bet

Thanks for posting, fun to sweat.

Cheers man, hopefully we'll make it to the end of this championship with a little extra pizza and beer money.

today
4-2 +4.06u on 8.4u risked
total
23-20 +2.42u on 60.13u risked

I know very little about snooker and I barely follow all the model talk in here, but I'm 100% loving the posts and updates. Thanks for the entertainment, rabbit!

You and me both. I've enjoyed what little snooker I've seen in the past, I think the players are really talented, but between the day job and everything else I haven't seen a single match I've bet on. It'll be interesting to see if someone who doesn't know or watch the sport can take a pure quantitative approach like this and grind out a decent return on investment.

today
1-2 -1.22u on 5.2u risked
total
24-22 +1.20u on 65.23u risked

Snooker numbers for Sunday:
akani v pengfei +-118, akani +127
dott v sharav -+324, no bet
holt v jones -+113, holt -102
li v higginson -+190, li -171
davis v carrington +-185, carrington -142
liang v lawler -+437, lawler +519
ning v robertson -+123, no bet
perry v clarke -+198, no bet
joyce v hamilton +-156, hamilton -119
walden v lines -+419, lines +623

I'll be surprised if either of the +519 or +623 plays actually hit tomorrow but, assuming I priced the actual win probabilities of Lawler and Lines with reasonable accuracy, I still took the +EV.

Just made the mistake of checking the live scores before heading into work and it's not looking very pretty right now hah.

One thing that has caught my attention over the past few days is when the model disagrees with the market on who the favorite is. For this morning I had Holt as the fave, the market had Jones as the fave, and Jones took it.

This has happened a few times recently, and it seems like the market is right enough in these situations that I shouldn't be disrespecting it. When I re-coded the Elo model to work within a frames based & best of n matchups environment, I took a few departures from the original implementation, and I'm going to look closer at these situations where it disagrees with the market to see if I've introduced a bug that is ultimately costing us units...

I just want to hop in and say how much I love this. It's exactly the right approach. Pick a niche'ish sport. Work on a model. Tweak it. Follow the market. Post plays. Make observations vs. the market based upon plays. Look into the specific subset that seems to be identified as bets. Is there something unique about that subset that the model identifies as a play that the market disagrees with? It could be a source of value or it could be something the model is missing. Track results. Try to watch a match or two with your newly trained eyes. Either way, it's being in the arena doing battle that generates feedback to improve upon.

Going through this process in a relatively public manner like this definitely encourages an acute focus and awareness that is hard to replicate in private settings. We've scraped, cleaned, analyzed and shared data. Tested a few different models for accuracy. Ran simulated wagering backtests to get an idea of potential ROI. And that all seemed like a "guaranteed to win" plan until days like today when I lose some of my own money and the model goes 6% into the red. In private I might have dismissed today as "negative variance", and maybe that's all it is, but now I'm looking at a few elements closer than I would have otherwise, so I don't mind that I paid 5.63 units for the lesson.

today
1-6 -5.63u on 8.34u risked
total
25-28 -4.43u on 73.57u risked

Has anyone been playing RSIF so far? I took a crack at trying to build a chart, and not sure if it flat out sucks or the variance is wild in RSIF. YTD I have RSIF yes @ 42.75% for the 130 games played so far. This by itself would suggest to me that it's just variance especially considering sample size, but I don't want to be totally naive and chase.

My data used to build a chart included 10,800 games with the historical results below.

To me intuitively it didn't make sense that a higher total game has a lower RSIF%. To counter this assumption I regressed the historical RSIF vs the total so that I could create RSIF numbers where the RSIF for my chart at 7.5 is greater than my RSIF number at a total of 7. Is this assumption disastrous, decent, or solid?

An additional assumption I'm making is that we are likely to be favoring the yes on RSIF pretty heavily, very rarely betting no RSIF based on the approach I'm taking. Books I've bet this usually have the no as a fav. Can anyone confirm from experience that I have this correct that yes bets are likely to outweight no bets by a considerable margin?

When using the RSIF numbers generated by the above process, taking any bet with edge > 0.00% based on my RSIF chart, the chart has an ROI of -21.12% n = 80. Filtering for only betting RSIF when edge is >2% , ROI is -20.51% n = 47. Filtering for only betting RSIF when edge is >3% , ROI is -7.11% n = 39 games.......so small sample sizes for my results and seems like RSIF been hitting far less than years past, but just hoping for a sanity check.

Snooker numbers for Monday. The first matchups start in about 3 hours but, unlike slaughterfest Sunday, the model was very close to market numbers for most matchups so I only found 3 plays.

brown v hallworth -+260, brown -235
gould v jones -+457, no bet
brecel v king -+192, no bet
odonnell v king -+101, no bet
xintong v filipiak -+553, no bet
vafaei v craigie -+102, no bet
wilson v greene -+193, greene +232
bingham v zifan -+1123, no bet
guodong v burns -+343, guodong -318
stevens v wakelin -+166, no bet

Per TomG's suggestions, I'll be looking closer at a few things over the next day or two and hopefully posting some worthwhile findings that generate more discussion.

I've been firing off a few 1st inning bets in the mornings before leaving for work. Down about 3 units right now; not taking it very seriously because I'm not able to look for plays on every game every day.


I agree with your intuition, and I would attribute the unexpected leaps in the opposite directions to the different sample sizes available for each total. And if I remember correctly from when I last analyzed that data set, the observed percentages for totals of 8 & 8.5 aren't even statistically significant given the sample size, meaning the true yes/no percentage might not be what is shown in that chart. I think trying to draw a line through the data via regression is at the least a decent idea.

Most of the plays I've found so far this season have been on the Yes. At the book I work at, I've been taking bets mostly on the no. I'm not a trader so can't see where all the money that comes in is going, but based on where the cash money I touch has been going, which is towards the No, I assume the lines are intentionally being shaded towards the no.

Sorry to hear about the ROI being so low. Just for clarity, are you flat or Kelly betting the RSIF? I remember sigley saying he always made money on RSIF, but he was Kelly betting it. I've been flat betting it fwiw.

dont use april and sept total data
markets not efficient

i set this up using my proj totals and RSIF/HRE odds based on modeling so people could beat BAS where i don't have an account

http://motorsportsodds.com/riptony/

RSIF is something like 21% ROI, HRE 19% to start the year.

BM killed the market tho along with MLB 2H.

Flat betting not kelly. I haven't bothered to see what the results would look like using Kelly yet. I'll take a look. I assumed from the start that I don't really know what I'm doing so flat betting was probably more prudent.

Sigley, I'd assume april data is not efficient due to sample size or projections just have wider error bars??? But why do you include September as well? Teams mailing it in at the end of the season?

projections, people don't know much about the teams, etc.

same with sept more people being called up they don't know, players resting etc.

see the same thing in all sports. nba first month totals are moving by 10 points everyday. then they were pretty stable from opening to close (unless someone was scratched) until the last couple of weeks when all the nba players stopped caring and are just killing time until the playoffs.

i took these draft picks remember when we crushed it last year trey lence o6.5 +105, sewell u5.5 -105, pitts o5.5 -125, waddle o11.5 -120, ja'maar o6 +130, devota o11.5 +102

Feels like such an odd year. Hard to pinpoint some things right now. I'm just glad I took Chase preseason +125 to be first WR and Najee to be first RB at one point when he dipped to plus odds.

I took a few last week that have all increased in odds.

- CB's O4 -175
- WR's O4.5 - 180

I actually took Chase U6 at -130 because while Pitts could come into play, the Bengals/Dolphins picks seems highly likely to be some combo of Penei/Chase. Feels like Pitts is the only one who could ruin that. Felt like Chase U6 at least pushes. I'm seeing Chase now at -225 to go U6, Penei -160 O5.5, Pitts -110 U5.5.

Jaycee Horn to be first defensive player has bounced around a little. +350 now. Had rose to +400. Not sure Parsons goes top 10 after sitting out a year. Surtain and Horn could each have suitors in the top 11.

I went back through my source code this morning, looked at a couple of the suspicious matchups, and it turned out that one of the coding "shortcuts" I had taken while in a hurry to get the model up and running caused some players to get assigned an unnaturally high Elo score. I wouldn't call it a bug, just that an assumption I made when making some custom modifications to the algorithm improved the model in most situations, but hurt in a few.

As it stands now, I've finally been able to get the model to (slightly) beat the closing lines on all 3 metrics of prediction accuracy, log loss, and Brier score in an out-of-sample test set of the most recent 10,000 matchups (excluding today). Inside that same subset I ran a wagering simulation:
n=3,598
1907 wins, 1691 losses
+587.76 units on 5,573.66 units risked
ROI = +10.55%

So I guess the final test is to keep betting and see if the real wagering results align with the simulated wagering results.