The Betting Thread

[Ben]

Some geezer (mate of a mate) supplied us with a massive trends document for Cheltenham, which suffice to say I did great from. Anyway, the week kind of reignited my passion for the horses so I've wasted my evening learning Python and I'm currently scraping every jump race result since 1998 from Racing Post, in anticipation of discovering trends for every course/meet.

Shite tips incoming.

This has been great fun. I decided the way to go was rather than just predict the finishing positions, I would instead try to predict the running time of each horse. The data scraping took me a week or so and then I had another week of errors because I'm shit at Python still and also didn't really understand how regression models and machine learning work.

Anyway, I finally had a breakthrough on Sunday night and the back testing was successful so last night I thought I'd try to predict today's racing. Unfortunately I've only set it up for chase races using 2018-2023 data so far (partly because I kept running into self-induced errors using the full dataset, but also I think my CPU won't cope with it all together) and there's a whopping two chase races today and they're both at the clown course Hereford so we've got just two shitty Class 5s, so not the best start. That might be a blessing in disguise though because I've not cracked the input of racecards yet so it took a lot of manual work to enter today's 15 runners into the model.

Today's predictions:
14:30 - #3 wins in 248.8303s (2nd #6, 3rd #2)
15:30 - #2 wins in 333.093s (2nd #4, 3rd #1)

Actual result:
14:30 - #3 wins in 248.9s (2nd #5, 3rd #2; paid 3 places)
15:30 - #5 wins in 336.6s (2nd #4, 3rd #2; paid only 2 places)

14:30 returns a win at 15/2 and a place at 11/4.
15:30 returns a place at 3/1. The predicted time for the actual winner was 336.2674s so again, incredibly, pretty bang on. Just turned out to be a slower race than anticipated.

I haven't obviously just stumbled across a quick route to millions but this is so much fun. It seems like it might be a good predictor, I just need to find a way to input racecards that isn't 100% manual for this to be worth bothering with, it takes way too long as it is. I'm excited to see what is says for the National on Saturday. Although I feel this model might work better for flats (less random opportunities for falls, even more trend indicators like stall draw).

[Manc]

Does the model account for juiced up Irish runners?

[Boydy]

This has been great fun. I decided the way to go was rather than just predict the finishing positions, I would instead try to predict the running time of each horse. The data scraping took me a week or so and then I had another week of errors because I'm shit at Python still and also didn't really understand how regression models and machine learning work.

Anyway, I finally had a breakthrough on Sunday night and the back testing was successful so last night I thought I'd try to predict today's racing. Unfortunately I've only set it up for chase races using 2018-2023 data so far (partly because I kept running into self-induced errors using the full dataset, but also I think my CPU won't cope with it all together) and there's a whopping two chase races today and they're both at the clown course Hereford so we've got just two shitty Class 5s, so not the best start. That might be a blessing in disguise though because I've not cracked the input of racecards yet so it took a lot of manual work to enter today's 15 runners into the model.

Today's predictions:
14:30 - #3 wins in 248.8303s (2nd #6, 3rd #2)
15:30 - #2 wins in 333.093s (2nd #4, 3rd #1)

Actual result:
14:30 - #3 wins in 248.9s (2nd #5, 3rd #2; paid 3 places)
15:30 - #5 wins in 336.6s (2nd #4, 3rd #2; paid only 2 places)

14:30 returns a win at 15/2 and a place at 11/4.
15:30 returns a place at 3/1. The predicted time for the actual winner was 336.2674s so again, incredibly, pretty bang on. Just turned out to be a slower race than anticipated.

I haven't obviously just stumbled across a quick route to millions but this is so much fun. It seems like it might be a good predictor, I just need to find a way to input racecards that isn't 100% manual for this to be worth bothering with, it takes way too long as it is. I'm excited to see what is says for the National on Saturday. Although I feel this model might work better for flats (less random opportunities for falls, even more trend indicators like stall draw).

Care to share what you've got?

I've been wanting to do similar for a while but machine learning makes my head hurt.

[Lofty]

Build us a full sheet for Aintree, please.

[Ben]

Does the model account for juiced up Irish runners?

If they've run in Britain before then yes! If there's some fresh blood then unfortunately not. It wouldn't be difficult to build in Irish data but I think I'm limited by CPU. I'm still adding in GB data from previous years and that's the priority at the moment. Hopefully I don't hit a memory error.

Care to share what you've got?

I've been wanting to do similar for a while but machine learning makes my head hurt.

Sure. I'm not at home until tonight so I'll paste up the code then. There's a load of models in the 'sklearn' package. I tested a few and ended up using RandomForestRegressor (DecisionTreeRegressor was also a good one). Also import train_test_split also from 'sklearn' to test the accuracy (R-squared?) score for your model. And of course 'pandas' to read the csv data input.

Build us a full sheet for Aintree, please.

Will do. My model accuracy score yesterday afternoon (for those Hereford predictions I posted) was 0.990 (max=1) and last night I added in another couple of years worth of data which brought the score up to 0.993. That is pretty huge. In the National (about 9 minutes run time), the 0.990 score leaves a margin of error of around 5 seconds; 0.993 shaves about 1.5 seconds off that margin of error, so if I ran the model on the race as it is, each horse would be within ~3.6s of the predicted time. Obviously the more I can bring it down the more accurate it will be, because even though 3.6s over 9 minutes is great, 3.6s can still cover a handful of positions. Given the number of years worth of data still missing from the model, I reckon that score can get >0.996 and then it could get tasty.

[Boydy]

Sure. I'm not at home until tonight so I'll paste up the code then. There's a load of models in the 'sklearn' package. I tested a few and ended up using RandomForestRegressor (DecisionTreeRegressor was also a good one). Also import train_test_split also from 'sklearn' to test the accuracy (R-squared?) score for your model. And of course 'pandas' to read the csv data input.

Yeah, it was sklearn I was looking at before. It was cleaning the data and all that shit beforehand I didn't know what to do with.

Had no idea what model I should be using or anything either. Did you just experiment with some?

[Ben]

Here's one for today from the model.

15:00 Market Rasen - War Lord (#1) has the fastest time but margin of error means a lot of overlap. Going off this I'd bet on it at 6/1 or more. It was 8/1 earlier, but drifted to 12/1 in the last hour.

16:00 Market Rasen - Coolmoyne (#9) was a good EW shout at 13/2 earlier (~50% shot for a top 3) but has been backed right in to 9/2F at the moment. The model has it as the winner but not that confident.

[Ben]

Yeah, it was sklearn I was looking at before. It was cleaning the data and all that shit beforehand I didn't know what to do with.

Had no idea what model I should be using or anything either. Did you just experiment with some?

Yep, loads of trial and error. I'm using RandomForestRegressor now but I went through loads to get there. There could easily be a better one out there still, but I just ran with it because it had a decent accuracy score right off the bat.

Data cleaning I feel your pain. Luckily my data scrapes in a somewhat logical fashion to begin with and as my Python skills still aren't great, I use Notepad++ to clean it up instead.

Cutting down my time spent on this is vital. Ideally it'd be ran on every race and the model would pick out a handful of value bets. As it is now, I'm inputting a limited number of races so I'm falling into the trap of trying to see value in something that probably isn't there, as the pain of admitting there's no good bets to be had after wasting an hour is not what you want.

[Baz]

EW double, let’s have it.

[Lofty]

Choo choo.

[Ben]

War Lord out to 16s now.

Something must not be right in the stables for a drift like that.

[Ben]

Our boy there was out in front and got excited and fell.

Since a few of you followed in, I'll put the stats up for transparency. As you can see it was anyone's race and I mentioned War Lord because his odds were so much longer than his chance. The drift this morning might have been from some stable info. I watched the race and it was incredibly slow (looked like it was pissing down). Bet365 paid 3 places (#5, #7, #4).

I've also added the 16:00 race below too. Bet365 paying 4 places.

15:00
num		pred(s)		min		max		WIN%
1		295.7262	293.6725	297.7942	20.58%
2		299.1247	297.0474	301.2165	3.73%
3		297.0416	294.9788	299.1188	14.06%
4		296.8301	294.7688	298.9058	15.11%
5		296.4196	294.3611	298.4925	17.14%
6		296.6480	294.5880	298.7225	16.01%
7		297.1812	295.1174	299.2594	13.37%

16:00					
num		pred(s)		min		max		WIN%
1		266.4029	264.5529	268.2659	0.12%
2		264.2181	262.3833	266.0658	12.93%
3		264.2582	262.4231	266.1062	12.69%
4		267.2189	265.3632	269.0876	0.00%
5		264.4423	262.6059	266.2915	11.61%
6		262.9835	261.1572	264.8225	20.16%
7		278.7130	276.7775	280.6620	0.00%
8		281.3068	279.3533	283.2740	0.00%
9		262.7363	260.9117	264.5736	21.61%
10		262.8642	261.0388	264.7024	20.86%
11		267.8636	266.0034	269.7368	0.00%

[Baz]

I’ve been told Not So Sobers in the 16:35 at Market Rasen. Was 16s but 8s now so looking promising.

[Ben]

A solid EW.

[Ben]

@Boydy

Toggle Spoiler

import pandas as pd
from sklearn.ensemble import RandomForestRegressor
from sklearn.model_selection import train_test_split

df = pd.read_csv('data.csv', low_memory=False)
# drop the column you want the model to predict
X = df.drop(columns=['secs'])
# insert the dropped column as y
y = df['secs']
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)

model = RandomForestRegressor()
model.fit(X_train, y_train)

score = model.score(X_test, y_test)
score

model.fit(X.values, y.values)
predictions = model.predict([ [# insert data input here] ])
predictions

[Boydy]

Cheers.

What's your original dataset look like? And what's your cleaning process?

[Boydy]

Also, I found this a while back which might be useful - it's a Racing Post scraper: https://github.com/joenano/rpscrape

[Ben]

That's what I'm using.

I'm having to make everything numerical. Luckily the scraper has a unique ID for courses, jockeys, horses, owners etc., so the only converting I need to to is things like Going (assign each type of Going a static number so the model knows it's a variable).

[Baz]

Is this Mellcorp/Bazbot version 3?

[Jimmy Floyd]

Masters

Hideki 19/1
Zalatoris 40/1
C.Young 50/1

[Ben]

I'm liking Zalatoris for a top 5.

[Waffdon]

I get Cameron young and smith mixed up so backed them both EW with Delta Work and Coko Beach respectively in the National.

Also backed the fav in both on the nose with another pony at aintree. Plus all these at a couple quid:

Eckroat 200/1 £3ew
Si Woo Kim 67/1 £5 ew
Nick Taylor 200/1 £3 ew
Bhatia 90/1 £3 ew
Denny McCarthy 200/1 £2 ew
Hojgaard 200/1 £5 ew
Lampracht 500/1 £2 ew

[Boydy]

That's what I'm using.

I'm having to make everything numerical. Luckily the scraper has a unique ID for courses, jockeys, horses, owners etc., so the only converting I need to to is things like Going (assign each type of Going a static number so the model knows it's a variable).

I grabbed some data for flats last night and all the jockey, horses, owners etc are text.

[Ben]

Go into your "rpscrape" folder and in there somewhere there's a settings.toml file - change jockey to false and jockey_id to true, then the scraper will get you the unique ID instead of their name.

[Boydy]

Ah, nice. Cheers.

[Ben]

Still doing chase only.

13:45 Aintree

Toggle Spoiler

num	pred		min		max		WIN%
1	318.9453235	316.8966043	321.0072875	0.00%
2	319.0448103	316.9954521	321.1074175	0.00%
3	314.2452779	312.226749	316.2768564	50.56%
4	314.335812	312.3167016	316.3679758	49.44%
5	318.8265473	316.778591	320.8877434	0.00%

A two horse race apparently. #3 is 4/1 and #4 is 11/10F so both technically within range to bet on as they're splitting hairs. I'm on #3.

14:55 Aintree

Toggle Spoiler

num	pred		min		max		WIN%
1	403.1735214	400.5837693	405.7800162	2.63%
2	398.5278056	395.9678948	401.104266	31.89%
3	413.0779144	410.4245422	415.7484405	0.00%
4	405.5922376	402.9869491	408.2143692	0.00%
5	399.5597048	396.9931657	402.1428364	25.39%
6	398.4222057	395.8629733	400.9979834	32.55%
7	402.3927372	399.8080004	404.9941842	7.54%

#6 is the favourite here (Shishkin) which is my own pick for this race anyway. I got 3/1 this morning. #2 currently at 10/1, great value there so I'm on that too.

16:05 Aintree

Toggle Spoiler

num	pred		min		max		WIN%
1	361.8681354	359.5437051	364.2075931	0.53%
2	360.9506345	358.6320976	363.2841606	4.59%
3	357.3537595	355.0583268	359.664032	20.48%
4	368.0078002	365.6439322	370.3869505	0.00%
5	365.0159145	362.6712646	367.3757224	0.00%
6	360.0065107	357.6940384	362.3339331	8.76%
7	360.13138	357.8181055	362.4596096	8.21%
8	366.5636131	364.2090217	368.9334268	0.00%
9	359.529654	357.2202447	361.8539935	10.86%
10	365.4473379	363.0999168	367.8099349	0.00%
11	364.2069227	361.8674693	366.5615005	0.00%
12	357.6479656	355.3506431	359.9601401	19.18%
13	358.5243889	356.2214368	360.8422295	15.31%
14	364.9912479	362.6467564	367.3508963	0.00%
15	360.6309023	358.3144192	362.9623613	6.00%
16	377.0275139	374.6057085	379.4649762	0.00%
17	364.4870099	362.1457574	366.8433985	0.00%
18	373.5829117	371.1832324	375.9981048	0.00%
19	361.0257751	358.7067555	363.3597869	4.26%
20	367.3907344	365.0308301	369.7658954	0.00%
21	361.5741777	359.2516356	363.911735	1.83%
22	365.7874632	363.4378573	368.1522591	0.00%

22 runners whittled down to half a dozen. #3 currently 40/1 and Bet365 paying 5 places.

16:40 Aintree

Toggle Spoiler

num	pred		min		max		WIN%
1	245.2348073	243.6595616	246.8202369	0.00%
2	244.6497904	243.0783025	246.2314379	0.00%
3	244.098749	242.5308006	245.676834	2.26%
4	244.8979661	243.324884	246.481218	0.00%
5	244.6303444	243.0589814	246.2118662	0.00%
6	250.0718257	248.4655098	251.6885264	0.00%
7	242.5611962	241.0031242	244.1293411	15.78%
8	241.9733979	240.4191015	243.5377426	20.95%
9	241.2270482	239.677546	242.7865678	27.51%
10	243.6564184	242.0913113	245.2316438	6.15%
11	244.646378	243.074912	246.2280034	0.00%
12	243.1180102	241.5563616	244.6897548	10.89%
13	247.9355266	246.342933	249.5384162	0.00%
14	244.6616726	243.0901084	246.2433969	0.00%
15	242.4843146	240.9267365	244.0519624	16.46%

If the Chase course is indeed Heavy (which it seems to be running as, despite the Soft designation), then the prediction is this instead

num	pred		min		max		WIN%
1	251.7068085	250.0899904	253.3340793	9.26%
2	253.1242964	251.4983732	254.7607311	0.64%
3	253.4395242	251.8115761	255.0779968	0.00%
4	253.6280571	251.998898	255.2677486	0.00%
5	254.1028269	252.4706182	255.7455878	0.00%
6	259.6992151	258.0310584	261.3781563	0.00%
7	251.1389081	249.5257379	252.7625074	12.72%
8	250.6973474	249.0870135	252.318092	15.41%
9	250.1839865	248.5769501	251.8014123	18.53%
10	252.6626538	251.0396959	254.296104	3.45%
11	253.5887208	251.9598144	255.228158	0.00%
12	250.7913294	249.1803918	252.4126816	14.84%
13	255.9766231	254.3323782	257.631498	0.00%
14	252.3380542	250.7171814	253.9694059	5.42%
15	249.9862663	248.3805	251.6024138	19.73%

#9 at 6/1 is my bet here.

[Waffdon]

I’ve got a fiver on Grey Dawning, Sir Gino, Shishkin, Bob Olinger & It’s on the Line

EW Yankee on Annamix, Unexpected Party, Ahoy Senor, Langer Dan & Honkey Tonk Highway ( fancying the Skelton’s a bit too much apparently)

[Ben]

A test run on the hurdles.

14:20 Aintree

Toggle Spoiler

num	pred		min		max		WIN%
1	298.8479008	296.8924197	300.8162618	0.00%
2	268.9581739	267.1982731	270.7296662	0.00%
3	267.3825623	265.6329714	269.1436768	0.00%
5	263.3497793	261.6265765	265.084332	57.89%
6	264.2983304	262.5689208	266.0391307	42.11%

Pretty much reflecting the bookies market.

[Jimmy Floyd]

Has number 1 only got three legs?

[Ben]

Has number 1 only got three legs?

Never gone over hurdles before and bookies have it at 125/1.

[Ben]

Last one for today.

15:30 Aintree

Toggle Spoiler

num	pred		min		max		WIN%
1	316.1211111	314.0526044	318.203242	0.00%
2	315.86569	313.7988546	317.9461386	0.00%
3	312.5245235	310.4795507	314.5829655	29.68%
4	316.5014083	314.4304131	318.586044	0.00%
5	316.2121808	314.1430782	318.2949115	0.00%
6	311.4359083	309.3980587	313.4871801	40.36%
7	312.4966974	310.4519066	314.5549561	29.96%
8	315.9155109	313.8483496	317.9962877	0

Not sure why it rates #6 so much (14/1) but I'm on it. Hurts to go against Langer Dan after Cheltenham.

[Waffdon]

Someone needs to blow Willie Mullins’ yard up. getting boring now.

[Manc]

Does the model account for juiced up Irish runners?

Nope.

[Ben]

26 out of 34 runners in the National are Irish.

[Ben]

Nope.

Yeah that was suss as fuck. All the others ran to predicted pace and this cunt just hit the afterburners.

[Ben]

A test run on the hurdles.

14:20 Aintree

Toggle Spoiler

num	pred		min		max		WIN%
1	298.8479008	296.8924197	300.8162618	0.00%
2	268.9581739	267.1982731	270.7296662	0.00%
3	267.3825623	265.6329714	269.1436768	0.00%
5	263.3497793	261.6265765	265.084332	57.89%
6	264.2983304	262.5689208	266.0391307	42.11%

Pretty much reflecting the bookies market.

Got the full finishing order correct and within 0.6s of winning time.

Forecast payout was miserable, however.

[Manc]

The size of these fields.

[Waffdon]

Thought Ahoy Senor had that

[Ben]

These first few races seem very slow for just "Soft" ground.

Indeed the GoingStick has it the same as the National course which is officially Heavy.

[Ben]

What a finish.

[Lofty]

That was a mad race.

[Waffdon]

Yeah, that was special. Would be tops off if Ahoy Senor and Langer Dan got up on the line. EW will still pay I guess

[Manc]

Nice little preview for the National that. Engine trumps everything.

[Waffdon]

I’ve got a fiver on Grey Dawning, Sir Gino, Shishkin, Bob Olinger & It’s on the Line

EW Yankee on Annamix, Unexpected Party, Ahoy Senor, Langer Dan & Honkey Tonk Highway ( fancying the Skelton’s a bit too much apparently)

That Langer Dan drift to 19s was nice. Around £130 back off £10.20 if the last two Skelton runners place.

[Manc]

[Ben]

Back in the office today so not got my laptop with me. Turned out alright yesterday in the end. 2 wins (including a forecast) and 2 EW.

Going solo on today's. 13:45 Aintree #2 and #5 reverse forecast paying about 12/1.

@Boydy how you getting on with yours?

[Boydy]

Back in the office today so not got my laptop with me. Turned out alright yesterday in the end. 2 wins (including a forecast) and 2 EW.

Going solo on today's. 13:45 Aintree #2 and #5 reverse forecast paying about 12/1.

@Boydy how you getting on with yours?

Haven't had time to actually grab the data without text yet and have a look at it.

I was reading this yesterday: https://medium.com/@chyun55555/data-...n-98437e8d93cb
I was wondering about this bit:

However, label encoding must be used only in appropriate cases. For example, using label encoding in regression models could result in critical errors because regression models will identify 2 as a larger value than 1. This means that regression models will identify grape(2) to be more important than banana(1), which definitely is not true because fruits are not ordinal data in this case.

Do you think that numbers for the ids of the horses, jockeys, trainers etc is kinda wrong then? The ids aren't ranking them but the algorithm is probably gonna consider larger numbers more important?

[Ben]

I don't think so. I'm using a decision tree model rather than a linear one, so it wouldn't consider a higher ID more important than a lower one. My understanding of decision tree so far is that the decision trees will identify trends in the data for that ID (ie. this jockey_ID usually goes a bit faster than average at this course_ID).

I might be wrong because I still don't completely understand how it works, but I've found a few different githubs exploring various things using the same models and they are appearing to be using unique IDs in this same way.

I suppose a way to confirm it is to use GridSearchCV on your data to assess whether the unique IDs are having a negative effect (if at all). I might try this tonight.

EDIT: Just read a bit more. At worst the IDs will be useless to the model (but certainly not detrimental). But if the machine learns a trend between the ID and another input then it will use it, otherwise it is essentially ignored. So something like course_ID will certainly be useful to the model (because course trends are definitely a thing).

[Waffdon]

£20 on the McManus treble today.

Feel dirty for backing Jonbon

[Ben]

Dem Irish roids again.

Shame we didn't get the forecast but a win is a win. 6/4 looked a decent price after watching him fly in the last furlong.