Exploring Value Betting in Football Match Odds – Exit Strategies

[TupleVision]

I now have a pretty rudimentary in-play model that said there was a little bit of value to lay the draw at half time in the Spurs game. Question to Kai, WtFH, why have you both chosen halftime to enter, have you discovered this is the optimal time, why not the 60th minute, 80th etc? Does having some second half stats not improve the accuracy rather than relying on one half where a team can come out and play a totally different game in the second half?

Secondly the whole exiting thing, do you just simply always leave it as a value bet, the Spurs game could've finished anything from 1-0, 4-4, 7-0 or 1-6 I left the bet in as a value bet but did have to sit on my hands a little. I'm assuming if you have a true edge in the market the variance plays out over time and not exiting is sub-optimal for PnL?

-------

The Spurs v Man Utd game, for me, didn't indicate any value pre-off, so left it alone, (My ZIP model odds tallied with the market odds). My approach is to focus on games that have value pre-off, and if the first half supports my pre-off findings, then make a decision for the 2nd half.


Tott v ManU.png

---------

The ZIP sheet, indicated correct scores, 1-1, 2-1, 1-0 with a squeak of an AOH. It was leaning towards more than one goal as well.


I've traded horse racing and greyhounds before, but with football, some games have a certain predictability about them, so letting the bet(s) run, realises the PnL Reckon it'll depend on how many games you get involved with, if it's every game, in every league, then trading out for green/red might be best. I'm only looking at the English & Scottish Prem for now, next year, will expand into the bigger euro leagues.


ZIP SpursvMUtd.png

---------

Looking at the final stats. (i didn't watch the game), it looked quite lively, but my New xG mirrored the Sofa-Score xG. That's a good sign, as I can lean on this sheet to support 2nd half decision making.


FT.png

Ah yes this makes sense, I now see what you're saying, combine YOUR pre-match assessment with halftime odds. I wonder if I can do this with my original and my new models in that case, thanks again.

[Kai]

Question to Kai, WtFH, why have you both chosen halftime to enter, have you discovered this is the optimal time, why not the 60th minute, 80th etc?

Not sure where I said that? You may be mixing things up with all the good feedback you've been getting from everyone

Does having some second half stats not improve the accuracy rather than relying on one half where a team can come out and play a totally different game in the second half?

Yeah, exactly. Liverpool-Wolves earlier was a good example, they finished the entire 2nd half with 0 xG and 0 shots which is not something you would expect at Anfield. That whole half started off on the wrong foot soon as they took Konate off for fear of a 2nd yellow, and Wolves being wolves smelt that bit of fear and weakness and rightfully went for it with nothing to lose. Every sub Slot made afterwards just weakened his control and grip over the match, and when he brought on Endo so early (which he never does) it summed up the pressure he felt his team were under (Carra remarked on this live as well), which basically hinted how the last 20ish mins were likely to play out. Not seen it yet but think Slot said something similar post-match to what I mentioned in the footy thread, the disallowed goals had a negative impact on the players mentally. But summarizing things after a match isn't difficult, forming your opinions during one is far more useful! Whether that's based on what you see unfolding or whatever your stats may be suggesting

[TupleVision]

I now have a pretty rudimentary in-play model that said there was a little bit of value to lay the draw at half time in the Spurs game. Question to Kai, WtFH, why have you both chosen halftime to enter, have you discovered this is the optimal time, why not the 60th minute, 80th etc? Does having some second half stats not improve the accuracy rather than relying on one half where a team can come out and play a totally different game in the second half?

Secondly the whole exiting thing, do you just simply always leave it as a value bet, the Spurs game could've finished anything from 1-0, 4-4, 7-0 or 1-6 I left the bet in as a value bet but did have to sit on my hands a little. I'm assuming if you have a true edge in the market the variance plays out over time and not exiting is sub-optimal for PnL?

-------

The Spurs v Man Utd game, for me, didn't indicate any value pre-off, so left it alone, (My ZIP model odds tallied with the market odds). My approach is to focus on games that have value pre-off, and if the first half supports my pre-off findings, then make a decision for the 2nd half.


Tott v ManU.png

---------

The ZIP sheet, indicated correct scores, 1-1, 2-1, 1-0 with a squeak of an AOH. It was leaning towards more than one goal as well.


I've traded horse racing and greyhounds before, but with football, some games have a certain predictability about them, so letting the bet(s) run, realises the PnL Reckon it'll depend on how many games you get involved with, if it's every game, in every league, then trading out for green/red might be best. I'm only looking at the English & Scottish Prem for now, next year, will expand into the bigger euro leagues.


ZIP SpursvMUtd.png

---------

Looking at the final stats. (i didn't watch the game), it looked quite lively, but my New xG mirrored the Sofa-Score xG. That's a good sign, as I can lean on this sheet to support 2nd half decision making.


FT.png

Well done on the model vs realised btw, it's looking very good

[TupleVision]

Question to Kai, WtFH, why have you both chosen halftime to enter, have you discovered this is the optimal time, why not the 60th minute, 80th etc?

Not sure where I said that? You may be mixing things up with all the good feedback you've been getting from everyone

Does having some second half stats not improve the accuracy rather than relying on one half where a team can come out and play a totally different game in the second half?

Yeah, exactly. Liverpool-Wolves earlier was a good example, they finished the entire 2nd half with 0 xG and 0 shots which is not something you would expect at Anfield. That whole half started off on the wrong foot soon as they took Konate off for fear of a 2nd yellow, and Wolves being wolves smelt that bit of fear and weakness and rightfully went for it with nothing to lose. Every sub Slot made afterwards just weakened his control and grip over the match, and when he brought on Endo so early (which he never does) it summed up the pressure he felt his team were under (Carra remarked on this live as well), which basically hinted how the last 20ish mins were likely to play out. Not seen it yet but think Slot said something similar post-match to what I mentioned in the footy thread, the disallowed goals had a negative impact on the players mentally. But summarizing things after a match isn't difficult, forming your opinions during one is far more useful! Whether that's based on what you see unfolding or whatever your stats may be suggesting

Apologies yep for some reason I had it in my head that you both only entered at half time but I think I've confused this with general advice about the entry being more important than the exit. I must admit I did bottle getting involved in the Liverpool match at halftime, I had only just finished my basic in-play model and it did indicate a bit of value to lay Liverpool, having watched the first half I thought it probably needed 'tweaking' a little but watching the second half maybe not, I did not anticipate the turn around at all though.

[TupleVision]

Looking at the final stats. (i didn't watch the game), it looked quite lively, but my New xG mirrored the Sofa-Score xG. That's a good sign, as I can lean on this sheet to support 2nd half decision making.

As soon as Gary Neville said in commentary "There's no way this game will end 1-0" I knew it would end 1-0, it's a shame you can't reverse engineer him as a betting model

[TupleVision]

Question, would you expect to see fair odds like the attached or is there something wrong with my model and it needs refining? Specifically the away odds, generally should my model be a lot closer to the exchange prices? This was the Spurs game vs Man Utd 1-0 at HT

[wearthefoxhat]

Question, would you expect to see fair odds like the attached or is there something wrong with my model and it needs refining? Specifically the away odds, generally should my model be a lot closer to the exchange prices? This was the Spurs game vs Man Utd 1-0 at HT

If you're getting exactly the same odds output as the market, then there would be no value opportunity. You have to trust the metrics you put through your algo washing machine, and take action if the model indicates value.

Tottenham @ 1.25 (your fair odds model) v 1.54 (live market) suggests fair win value of 1.15 (15%). How you then stake is a personal choice. 1/4 kelly or a staking plan that stakes more if there's more value on offer.

eg:

Value Staking.png

------

If the fair odds were 1.15

Odds 1.15.png

[TupleVision]

Question, would you expect to see fair odds like the attached or is there something wrong with my model and it needs refining? Specifically the away odds, generally should my model be a lot closer to the exchange prices? This was the Spurs game vs Man Utd 1-0 at HT

If you're getting exactly the same odds output as the market, then there would be no value opportunity. You have to trust the metrics you put through your algo washing machine, and take action if the model indicates value.

Tottenham @ 1.25 (your fair odds model) v 1.54 (live market) suggests fair win value of 1.15 (15%). How you then stake is a personal choice. 1/4 kelly or a staking plan that stakes more if there's more value on offer.

eg:


Value Staking.png

------

If the fair odds were 1.15


Odds 1.15.png

Interesting my model's odds weren't too far for the home team compared to yours.

I'm concerned about the away team, this in-play is basically my pre-match model with time decay added and a function to allow for the current number of goals for each side. Pre-match my tissue odds were close to the exchange, more often than not lower than the exchange by a few ticks. When I found a match that had odds with a larger discrepancy I had confidence that was value. Never though did I have exchange odds of 8.0 and tissue of 16.0. I'm not sure if I've missed something with the time decay factor, I struggle to believe the exchange can be that wrong after 45mins?

[wearthefoxhat]

With the time decay, can you code a "speed up" function as the game starts?

eg: Minutes - First half 0 -20 slow, 21-45 steady, 45+break-70, moderate-fast, 71-85, fast, 86-90added, very fast.

With the away odds, they are just part of the overall 100% book. If your metrics are leaning towards the home team, then the draw/away will be extended by %. If the game remains 0-0, despite the metrics leaning towards the home side, then the draw odds will decay normally up until half-time. The away odds will extend.

The attack/momentum graphs along with the supporting stats do paint a fair picture without actually watching the game, and it does, to some extent, replicate the market too.

[TupleVision]

With the time decay, can you code a "speed up" function as the game starts?

eg: Minutes - First half 0 -20 slow, 21-45 steady, 45+break-70, moderate-fast, 71-85, fast, 86-90added, very fast.

With the away odds, they are just part of the overall 100% book. If your metrics are leaning towards the home team, then the draw/away will be extended by %. If the game remains 0-0, despite the metrics leaning towards the home side, then the draw odds will decay normally up until half-time. The away odds will extend.

The attack/momentum graphs along with the supporting stats do paint a fair picture without actually watching the game, and it does, to some extent, replicate the market too.

This is a good point. For a horrifying moment there I thought the probability didn't add up to 100%. It's 99.55% which I'll take.

I see what you're saying about the extended % and no my model doesn't currently allow for that type of speed change, ChatGPT is going to be worked hard today

[ShaunWhite]

With the time decay, can you code a "speed up" function as the game starts?

eg: Minutes - First half 0 -20 slow, 21-45 steady, 45+break-70, moderate-fast, 71-85, fast, 86-90added, very fast.

With the away odds, they are just part of the overall 100% book. If your metrics are leaning towards the home team, then the draw/away will be extended by %. If the game remains 0-0, despite the metrics leaning towards the home side, then the draw odds will decay normally up until half-time. The away odds will extend.

The attack/momentum graphs along with the supporting stats do paint a fair picture without actually watching the game, and it does, to some extent, replicate the market too.

This is a good point. For a horrifying moment there I thought the probability didn't add up to 100%. It's 99.55% which I'll take.

I see what you're saying about the extended % and no my model doesn't currently allow for that type of speed change, ChatGPT is going to be worked hard today

Don't forget to add a margin to the fair price calc. At 100% book your 'perfect' model will break even.

If you want to look at the rate of change (the speed up function), then you want to use calculus,

To save my typing the gpt explanation is....
Calculus is a branch of mathematics that deals with change and rates of change. It has two main parts:

1. Differentiation – Focuses on rates of change (e.g., how fast odds are changing over time).
2. Integration – Focuses on accumulating quantities (e.g., total price movement over a period).

It’s useful when you need to understand how something is changing at any given moment or over time—like your price decay acceleration example.

I did a calculus course on Udemy because it's over 40yrs since I failed my 'A' - level maths

[Fugazi]

With the time decay, can you code a "speed up" function as the game starts?

eg: Minutes - First half 0 -20 slow, 21-45 steady, 45+break-70, moderate-fast, 71-85, fast, 86-90added, very fast.

With the away odds, they are just part of the overall 100% book. If your metrics are leaning towards the home team, then the draw/away will be extended by %. If the game remains 0-0, despite the metrics leaning towards the home side, then the draw odds will decay normally up until half-time. The away odds will extend.

The attack/momentum graphs along with the supporting stats do paint a fair picture without actually watching the game, and it does, to some extent, replicate the market too.

This is a good point. For a horrifying moment there I thought the probability didn't add up to 100%. It's 99.55% which I'll take.

I see what you're saying about the extended % and no my model doesn't currently allow for that type of speed change, ChatGPT is going to be worked hard today

You might be better of modelling something more obscure if youre doing it with gpt. Football, especially at the top level has so many people doing this youll struggle to beat them. Too many factors needing a lot of time and computing power. For example the weather, distance travelled to the game, how long since last game etc etc. The factors going into the best models is never ending and needs a team constantly working on it.

Where this can work is when you really study a lower league team, follow team news etc etc. You then combine this knowledge with a model.... I would save yourself the time and use a model that already exists. You'll need to be tracking how different combinations of line ups impact expected results.

If you live near a conference team l, for example, go watch it live and you can coin decent ROI in if you dont mind standing in the cold. You'll want to optimise your set up to be able to quickly place bets. Even better if you follow the team for a long time you find ways to get inside knowledge on injuries etc pre match.

[henbet22]

Question to Kai, WtFH, why have you both chosen halftime to enter, have you discovered this is the optimal time, why not the 60th minute, 80th etc?

Not sure where I said that? You may be mixing things up with all the good feedback you've been getting from everyone

Does having some second half stats not improve the accuracy rather than relying on one half where a team can come out and play a totally different game in the second half?

Yeah, exactly. Liverpool-Wolves earlier was a good example, they finished the entire 2nd half with 0 xG and 0 shots which is not something you would expect at Anfield. That whole half started off on the wrong foot soon as they took Konate off for fear of a 2nd yellow, and Wolves being wolves smelt that bit of fear and weakness and rightfully went for it with nothing to lose. Every sub Slot made afterwards just weakened his control and grip over the match, and when he brought on Endo so early (which he never does) it summed up the pressure he felt his team were under (Carra remarked on this live as well), which basically hinted how the last 20ish mins were likely to play out. Not seen it yet but think Slot said something similar post-match to what I mentioned in the footy thread, the disallowed goals had a negative impact on the players mentally. But summarizing things after a match isn't difficult, forming your opinions during one is far more useful! Whether that's based on what you see unfolding or whatever your stats may be suggesting

Game of two halves innit?
Used to be hitting the bar/post/big save now its Var disallowing goals that should prick your ears up.
Dunno how you add that to a model though

[TupleVision]

With the time decay, can you code a "speed up" function as the game starts?

eg: Minutes - First half 0 -20 slow, 21-45 steady, 45+break-70, moderate-fast, 71-85, fast, 86-90added, very fast.

With the away odds, they are just part of the overall 100% book. If your metrics are leaning towards the home team, then the draw/away will be extended by %. If the game remains 0-0, despite the metrics leaning towards the home side, then the draw odds will decay normally up until half-time. The away odds will extend.

The attack/momentum graphs along with the supporting stats do paint a fair picture without actually watching the game, and it does, to some extent, replicate the market too.

This is a good point. For a horrifying moment there I thought the probability didn't add up to 100%. It's 99.55% which I'll take.

I see what you're saying about the extended % and no my model doesn't currently allow for that type of speed change, ChatGPT is going to be worked hard today

You might be better of modelling something more obscure if youre doing it with gpt. Football, especially at the top level has so many people doing this youll struggle to beat them. Too many factors needing a lot of time and computing power. For example the weather, distance travelled to the game, how long since last game etc etc. The factors going into the best models is never ending and needs a team constantly working on it.

Where this can work is when you really study a lower league team, follow team news etc etc. You then combine this knowledge with a model.... I would save yourself the time and use a model that already exists. You'll need to be tracking how different combinations of line ups impact expected results.

If you live near a conference team l, for example, go watch it live and you can coin decent ROI in if you dont mind standing in the cold. You'll want to optimise your set up to be able to quickly place bets. Even better if you follow the team for a long time you find ways to get inside knowledge on injuries etc pre match.

You make some valid points—top-tier football markets are highly efficient, and lower leagues can offer more opportunities due to less sophisticated modeling. Live betting and insider knowledge can definitely provide an edge.

However, dismissing statistical modeling entirely isn’t accurate. While GPT alone isn’t ideal for this, well-designed models (like Poisson-based or zero-inflated approaches) can still find value, especially in niche markets like halftime betting or goal totals. The key is focusing on inefficiencies rather than competing directly with sportsbooks’ full-market models.

[TupleVision]

Don't forget to add a margin to the fair price calc. At 100% book your 'perfect' model will break even.

Sorry Shaun 'my' margin or the bookies margin? I'm not sure I totally understand this statement.