| Your help needed 12:30 - May 7 with 1543 views | backinbeige |
I’ve spent the last short while making a spreadsheet to predict the results of football matches. I’ll let you guess why.
It’s reasonably successful — it’s quite good at predicting value bets. I tend to stake £12 each weekend and am £100 up roughly every six months. I’ll take that. (If you had guessed betting, you would be correct.)
It conducts four tests on the upcoming fixtures based on this season’s scores so far. For each match it looks at:
1. Goals scored vs goals conceded for both teams
Calculates how many goals the home team has scored in the last five matches vs the league average for that period, and how many the away team has conceded vs the average. If the home team is better by a buffer of 1 then home win predicted. The same in reverse for the away team but with a 1.2 buffer (as it’s harder to win away).
2. Points per game last 30 days
How many points the home and away team are averaging per match in the last month. If one team is outperforming the other by a buffer of 1 or 1.2 then recommend that win.
3. Form score (who they actually earned their points against)
Works out how many points a team has earned were actually ‘worth’ based on their opponents PPG for the previous 30 days. The points earned are multiplied against the other team’s PPG. For example, if Manchester City earned three points against Fulham who were averaging 1.5 PPG until that point then Manchester City’s form score for that game is 4.5 (3 points earned X Fulham’s 1.5 PPG). However, if Fulham had won and Manchester City had an average PPG of 3 until that point then Fulham’s Form Score would be 9 (3 points earned X City’s 3 PPG) for that game.
For an upcoming fixture, an average is taken of each team’s Form Score for the last 5 games to work out how well they’re doing. If they’re winning, and if they’re winning against decent teams. The two averaged Form Scores are compared and a result recommended if a buffer is exceeded.
4. Performance against percentile
Last year’s league table is split into five percentiles (0-20%, 20-40%, etc). Works out how many PPG a team earned last year against the teams in the same league percentile as an opponent. For example, if Aston Villa inexplicably dropped lots of points against teams in the relegation zone last season but seemed to get results against teams in the second top percentile, and they are up against a team who finished in the two top percentile, the spreadsheet will flag it.
It usually says a team that finished well last season will beat a team that didn’t, but it also picks up on those anomalies where a team seemed to sneak something against the top teams and comes unstuck against the bottom teams. Perhaps because they are a defensive minded team that does well when it shuts up shop against a decent team but comes unstuck when it tries to play attacking football against a team it thought it would beat.
*
So that’s the spreadsheet.
My question to you is:
What are the other tests I can include?
What are the other situations that cause you to think a team is likely to beat another one? It doesn’t have to be a fully constructed test, I’m happy to waste hours of my life trying to figure that out. Instead, I’m keen to find out what you think are those obvious indicators that a team is going to win, based on past performance.
The spreadsheet isn’t perfect. It’s really good at predicting teams who should win and what would happen if everything goes according to plan, but rarely happens. It gets completely caught out by the freak results, or just the slightly unexpected results.
A couple of tests I’m thinking of adding are:
5. Comparing previous predictions
Some kind of rolling score looking at the prediction of the spreadsheet vs the actual outcome. An average of how many times the prediction was correct can be used to determine its accuracy.
The accuracy of all four (or more) tests can be used to give a weighting as to how much consideration I should take into account for all outcomes.
For example, the four test is probably the least accurate as it relies on last year’s table. Teams change year on year. Players change, managers change. So, it’s a useful test but only to a point.
6. Combine the weighting of each test into a score and compare against odds
If test 1 is correct 80% of the time and test two is correct 20% of the time, and test 1 is predicting a win, then it’s 80% likely. That’s 1/4 on, or 1.25 as a decimal. If the odds are significantly longer than that then it’s worth investigating further.
Please let me know any tests you reckon should be done, or you do in your mind before a game, based on past results only (not ‘new manager bounce’ unfortunately) that I should consider doing.
In answer to your next question: Reasonably fun at parties, but admittedly not always.
876 words including this sentence, if you’re wondering. [Post edited 7 May 2023 14:36]
|  |
| RVRB: https://app.rvrb.one/channels/have-guitar-will-listen -
Listen to music when WFH, set your own music to DJ and find new recommendations from others |
| |  |
| Your help needed on 12:38 - May 7 with 1477 views | Kropotkin123 |
Home and away weighting? |  |
| |
| Your help needed on 11:12 - May 9 with 1348 views | backinbeige |
| Your help needed on 12:38 - May 7 by Kropotkin123 |
Home and away weighting? |
Sounds good - as in working out how well a team is currently doing home or away? | |
| RVRB: https://app.rvrb.one/channels/have-guitar-will-listen -
Listen to music when WFH, set your own music to DJ and find new recommendations from others |
| |
| Your help needed on 11:22 - May 9 with 1319 views | Skip_Intro |
Are you up £100 every 6 months or is that a typo?
That's a return of £100 for an outlay of just over £300...
If it's not a typo wouldn't you'd be better off backing a couple of odds on favourites with bigger stakes every 6 months and saving yourself hours and hours of work?
I'm no gambling expert - just looking at the numbers by the way. |  | | |
| Your help needed on 11:42 - May 9 with 1277 views | backinbeige |
| Your help needed on 11:22 - May 9 by Skip_Intro |
Are you up £100 every 6 months or is that a typo?
That's a return of £100 for an outlay of just over £300...
If it's not a typo wouldn't you'd be better off backing a couple of odds on favourites with bigger stakes every 6 months and saving yourself hours and hours of work?
I'm no gambling expert - just looking at the numbers by the way. |
True it's not great but it is a profit - especially when most people make a loss at the bookies. https://www.theguardian.com/business/2023/jan/06/bet365-boss-denise-coates-was-p It's also small bets mostly to prove the spreadsheet works, which should be more of a guaranteed bet than hosing money at long odds every six months. [Post edited 9 May 2023 13:55]
| |
| RVRB: https://app.rvrb.one/channels/have-guitar-will-listen -
Listen to music when WFH, set your own music to DJ and find new recommendations from others |
| |
| Your help needed on 12:14 - May 9 with 1219 views | Stenvict |
What about head to heads? Some teams have certain teams that they rarely beat despite being on paper significantly better. For example, we've never beaten Cheltenham in the league, but on paper we should beat them. |  |
| |
| Your help needed on 12:34 - May 9 with 1180 views | JakeITFC |
It's definitely a good starting point (and an impressive ROI, so well done). Your next step I think will be going into the events that happened in games below the surface level (i.e. doing legwork that other people don't want to do). That's where the xG bods etc. will be making their assessments (and why Ipswich were probably shining bright red in your model in that spell we kept drawing post-WC but were still 1/4 or similar) and so maybe something to be said for going the traditional way. I would also say that carving out a specialism is a worthwhile thing to do - the Premier League is the most watched market in the world so it's hard to find an edge, but you probably know more about Ipswich U23 games than any model could just by reading this board (so try and find avenues to information that isn't readily available in other places, admittedly this likely to be more qualitative than quantitative).
I think a lot of what you are doing is going to be similar to a lot of other models (including bookies setting the opening line) so your edge is going to be improved by accessing markets early and having a large range of bookmakers to bet with, so that may be another good dataset to add to your modelling. |  | | |
| Your help needed on 14:08 - May 12 with 936 views | backinbeige |
| Your help needed on 12:34 - May 9 by JakeITFC |
It's definitely a good starting point (and an impressive ROI, so well done). Your next step I think will be going into the events that happened in games below the surface level (i.e. doing legwork that other people don't want to do). That's where the xG bods etc. will be making their assessments (and why Ipswich were probably shining bright red in your model in that spell we kept drawing post-WC but were still 1/4 or similar) and so maybe something to be said for going the traditional way. I would also say that carving out a specialism is a worthwhile thing to do - the Premier League is the most watched market in the world so it's hard to find an edge, but you probably know more about Ipswich U23 games than any model could just by reading this board (so try and find avenues to information that isn't readily available in other places, admittedly this likely to be more qualitative than quantitative).
I think a lot of what you are doing is going to be similar to a lot of other models (including bookies setting the opening line) so your edge is going to be improved by accessing markets early and having a large range of bookmakers to bet with, so that may be another good dataset to add to your modelling. |
Thank you!
You're right the next step is trying to incorporate the odds and weighing my predictions against the market to see if something is undervalued. This way I have to be wrong less times, as when I'm correct I'll win bigger. Betfair seem to have downloadable odds data to input into spreadsheets, correct to the minute.
The other - lazy, sorry - caveat I have is I'm trying to cut down on my copy/paste admin. There are some websites out there that give away spreadsheets of the most recent results each week, meaning I don't have to type out every team name and input every result. I just copy/paste the results in. If I'm to delve into player or in-game stats I'd need to find a similar spreadsheet that someone else has put together, so I can just paste the stats in, rather than type them out for the 92 English teams and 42 Scottish teams. Life is for living. | |
| RVRB: https://app.rvrb.one/channels/have-guitar-will-listen -
Listen to music when WFH, set your own music to DJ and find new recommendations from others |
| |
| Your help needed on 15:05 - May 12 with 863 views | StokieBlue |
Nice explanation of your tool, from what you've written I think it'll mostly predict heavy favourites in line with the bookies when in reality you want to try and isolate the bets which go against the bookies.
This is obviously an extremely hard thing to achieve.
You're essentially using a RWA (rolling weighted average) approach to ensure your data is not stale (ie. within 30 days) which is sensible and then generating attacking and defending scores which you than feed through your process to translate into a result.
You could implement a Poisson model using those attacking and defending scores you've already generated to get a distribution of probabilities for various scorelines. You could then sum up the probabilities to get a more "accurate" weighting to apply to the scores.
You could also look at the number of shots a team has versus how many shots the other team concedes (weighted for home and away). You can get more granular data at places like this:
https://www.football-data.co.uk/englandm.php
Your inputs would also seem to lend themselves to some ML, perhaps a random forest approach but be warned it'll be very difficult to get a model well enough trained to beat the bookies.
All of the above is about predicting a scoreline though - a winner and a loser and the bookies are very, very good at that. Have you considered trying to predict something more obscure?
For instance, odds for bets such as half-time, full-time are usually quite generous. If you could come up with a process which predicts when a "lesser" team might take the lead against a stronger opponent and hold on until half-time and then likely lose the second half (ie. your current full-time prediction) you'll have to get a lot less successful bets in order to make a profit.
Once again that's easier said than done but you can look at factors like:
- Whether the weaker team is likely to score at all.
- Whether they score more in the first or second half.
- Whether they score more from set-pieces (ie. good chance of sneaking a 1st half goal).
- Their historic half-time results.
Trying to beat the bookies at their own game on full-time results maybe fun but it's unlikely to be a long-term winning strategy, hence looking for outside bets where less analysis may have been performed.
SB |  |
| Avatar - IC410 - Tadpoles Nebula |
| |
| |
By continuing to use the site, you agree to our use of cookies and to abide by our Terms and Conditions.
