False edge? Am i fooled by variance?

Post Reply
offlimit88
Posts: 68
Joined: Mon Feb 25, 2019 2:29 pm

Immagine.png
Hi, sorry for my english,i would like your opinion on this.
I am trying a new idea, i trade basically on stats. The problem with stats is that 1) Everyone knows them on average 2) You can 't be sure if stats will be similar in the future.
In this case i used some stats of 2018 and backtested on 3 months of 2019.

After 120 events, the market sets the probabilites of this event to occur in average 32 % of the time.
Based on my 1400 events research, my selections should give me a strike rate of 37%, so i'm quite sure i am into something.
Then why am i currently having a strike rate of 48%?

I developed a lot of strategies, those that start very promising usually even up at a certain point. What i am expecting here, is that the curve will begin to drop brutally, in order to match the probabilities. I want to clarify that my study is correct.

Let's say it is just good luck, did this ever happen to you? Could it be that there is something more in my selections that i'm currently not understanding?
You do not have the required permissions to view the files attached to this post.
User avatar
murdok
Posts: 151
Joined: Sun Apr 02, 2017 7:10 pm

if you have more events the date will be more accurate
eightbo
Posts: 2154
Joined: Sun May 17, 2015 8:19 pm
Location: Malta / Australia

you only have ~100 mkts of actual results, it's not enough
If you flip a coin 100 times it would not be so uncommon to get a 60/40 split, so 60% when we know it should be 50% — there's your 10% difference you are experiencing. ...but by the time you reach 1,000 you will be far closer to expected probability (reality of your edge)

just continue doing things as you are unchanged to more accurately answer your question of if you were or weren't fooled by randomness (variance)
https://en.wikipedia.org/wiki/Law_of_large_numbers
offlimit88
Posts: 68
Joined: Mon Feb 25, 2019 2:29 pm

eightbo wrote:
Sat Jun 27, 2020 5:18 pm
you only have ~100 mkts of actual results, it's not enough
If you flip a coin 100 times it would not be so uncommon to get a 60/40 split, so 60% when we know it should be 50% — there's your 10% difference you are experiencing. ...but by the time you reach 1,000 you will be far closer to expected probability (reality of your edge)

just continue doing things as you are unchanged to more accurately answer your question of if you were or weren't fooled by randomness (variance)
https://en.wikipedia.org/wiki/Law_of_large_numbers
Yes i know about the law of large numbers.
But in this case it s not a 60/40, a 10 % difference from a 50-50 strategy. It s actually a 50/50 as opposed to a 30/70, so it s 20%.
Jukebox
Posts: 1576
Joined: Thu Sep 06, 2012 8:07 pm

offlimit88 wrote:
Sat Jun 27, 2020 1:48 pm
Immagine.png
Hi, sorry for my english,i would like your opinion on this.
I am trying a new idea, i trade basically on stats. The problem with stats is that 1) Everyone knows them on average 2) You can 't be sure if stats will be similar in the future.
In this case i used some stats of 2018 and backtested on 3 months of 2019.

After 120 events, the market sets the probabilites of this event to occur in average 32 % of the time.
Based on my 1400 events research, my selections should give me a strike rate of 37%, so i'm quite sure i am into something.
Then why am i currently having a strike rate of 48%?

I developed a lot of strategies, those that start very promising usually even up at a certain point. What i am expecting here, is that the curve will begin to drop brutally, in order to match the probabilities. I want to clarify that my study is correct.

Let's say it is just good luck, did this ever happen to you? Could it be that there is something more in my selections that i'm currently not understanding?
How unusual was a run of 120 events like this in your stats?
rik
Posts: 1583
Joined: Sat Jan 25, 2014 5:16 am
Location: London

offlimit88 wrote:
Sat Jun 27, 2020 1:48 pm
Immagine.png
What i am expecting here, is that the curve will begin to drop brutally, in order to match the probabilities. I want to clarify that my study is correct.
why would it drop brutally, it should flatten if your estimation of 37% is accurate, that its overperformed if at all it should make your expectation for future results better not worse
This is pre race horse racing?
User avatar
jamesedwards
Posts: 2234
Joined: Wed Nov 21, 2018 6:16 pm

Here is my latest "Think-I've-Cracked-It' automation. :roll:

500 markets in it was all roses. 1,000 markets in and it's a complete dog!

latest.PNG
You do not have the required permissions to view the files attached to this post.
User avatar
ShaunWhite
Posts: 9731
Joined: Sat Sep 03, 2016 3:42 am

jameegray1 wrote:
Sun Jun 28, 2020 2:14 am
Here is my latest "Think-I've-Cracked-It' automation. :roll:

500 markets in it was all roses. 1,000 markets in and it's a complete dog!
Problem is James that charts are like fractals, you don't know what zoom level you're looking at. That dip might be almost invisble after a year, and you might not know what happened before it started. How does it look if you put a trendline on that? origin zero. I'm guessing it's still +ve.
offlimit88
Posts: 68
Joined: Mon Feb 25, 2019 2:29 pm

Jukebox wrote:
Sat Jun 27, 2020 11:11 pm
offlimit88 wrote:
Sat Jun 27, 2020 1:48 pm
Immagine.png
Hi, sorry for my english,i would like your opinion on this.
I am trying a new idea, i trade basically on stats. The problem with stats is that 1) Everyone knows them on average 2) You can 't be sure if stats will be similar in the future.
In this case i used some stats of 2018 and backtested on 3 months of 2019.

After 120 events, the market sets the probabilites of this event to occur in average 32 % of the time.
Based on my 1400 events research, my selections should give me a strike rate of 37%, so i'm quite sure i am into something.
Then why am i currently having a strike rate of 48%?

I developed a lot of strategies, those that start very promising usually even up at a certain point. What i am expecting here, is that the curve will begin to drop brutally, in order to match the probabilities. I want to clarify that my study is correct.

Let's say it is just good luck, did this ever happen to you? Could it be that there is something more in my selections that i'm currently not understanding?
How unusual was a run of 120 events like this in your stats?
That's curious. In those 3 months my study shows a strike rate of 38/39% in that period, circa 300 event.
The only thing i can t backtest is the actual market price, i usually play above a certain price. So i did not play all those 300 events, just these 120.
offlimit88
Posts: 68
Joined: Mon Feb 25, 2019 2:29 pm

jameegray1 wrote:
Sun Jun 28, 2020 2:14 am
Here is my latest "Think-I've-Cracked-It' automation. :roll:

500 markets in it was all roses. 1,000 markets in and it's a complete dog!


latest.PNG
Wow that's sick, is it tennis?
Anyway we are not talking about same percentages, as i understand from your picture. Did you backtest it?
If i take 200 of you events, there is always a major correction. I'm waiting for mine.

This one for instance, you can clearly see outperforming, and after 50 events, it comes the correction
Immagine3.png
rik wrote:
Sat Jun 27, 2020 11:22 pm

why would it drop brutally, it should flatten if your estimation of 37% is accurate, that its overperformed if at all it should make your expectation for future results better not worse
This is pre race horse racing?
Tennis. It shoud revert to the mean, but usually the faster you go up, the faster you go down, assuming it is outperformed.

First time i see a 30/70 strategy with this win ratio.
Immagine2.png
You do not have the required permissions to view the files attached to this post.
User avatar
jamesedwards
Posts: 2234
Joined: Wed Nov 21, 2018 6:16 pm

ShaunWhite wrote:
Sun Jun 28, 2020 5:32 am
jameegray1 wrote:
Sun Jun 28, 2020 2:14 am
Here is my latest "Think-I've-Cracked-It' automation. :roll:

500 markets in it was all roses. 1,000 markets in and it's a complete dog!
Problem is James that charts are like fractals, you don't know what zoom level you're looking at. That dip might be almost invisble after a year, and you might not know what happened before it started. How does it look if you put a trendline on that? origin zero. I'm guessing it's still +ve.
You are correct, it does. Perhaps I will try running it again once I am feeling a little richer ;)
latest v2.PNG
You do not have the required permissions to view the files attached to this post.
User avatar
jamesedwards
Posts: 2234
Joined: Wed Nov 21, 2018 6:16 pm

offlimit88 wrote:
Sun Jun 28, 2020 10:40 am

Wow that's sick, is it tennis?
No, not tennis actually. It just seemed a good example to share.
User avatar
ShaunWhite
Posts: 9731
Joined: Sat Sep 03, 2016 3:42 am

It's about understanding your edge. If you see an anomaly in the data then without a real world reason it's a prime candidate for reversion. And if you only trade +ve anomalies then you'll probably always just end up getting the worst of it.

It's like picking lottery numbers, are you the sort that chooses "6" because 6 has come out 10 times in a row, or do you choose "6" because it hasn't been out for 20 weeks straight. Personally when I see an anomaly I can't explain I oppose it rather than assuming the god of stats will keep rolling the same dice.
User avatar
Crazyskier
Posts: 1157
Joined: Sat Feb 06, 2016 6:36 pm

jameegray1 wrote:
Sun Jun 28, 2020 3:38 pm
ShaunWhite wrote:
Sun Jun 28, 2020 5:32 am
jameegray1 wrote:
Sun Jun 28, 2020 2:14 am
Here is my latest "Think-I've-Cracked-It' automation. :roll:

500 markets in it was all roses. 1,000 markets in and it's a complete dog!
Problem is James that charts are like fractals, you don't know what zoom level you're looking at. That dip might be almost invisble after a year, and you might not know what happened before it started. How does it look if you put a trendline on that? origin zero. I'm guessing it's still +ve.
You are correct, it does. Perhaps I will try running it again once I am feeling a little richer ;)

latest v2.PNG
Reversion to mean can indeed be a bitch!

I will only say that the Maria Strategy (layed horses from £1k to £100k successfully) lost many hundreds of pounds before it went on to make thousands, so patience and BIG data is key.

I spend many sleepless hours analysing each day's results, and actually think I'm on to something with the greyhounds and back odds ordering and selecting certain traps at certain courses, etc etc, but only time, and lots of it will tell. At least with dogs there are around 120 races daily to get fast input. I struggle with the horses and they don't have the bumping and barging that often sees very short-priced dogs overturned.

CS
User avatar
jamesedwards
Posts: 2234
Joined: Wed Nov 21, 2018 6:16 pm

ShaunWhite wrote:
Sun Jun 28, 2020 5:32 am

Problem is James that charts are like fractals, you don't know what zoom level you're looking at. That dip might be almost invisble after a year, and you might not know what happened before it started. How does it look if you put a trendline on that? origin zero. I'm guessing it's still +ve.
Crazyskier wrote:
Sun Jun 28, 2020 7:28 pm

Reversion to mean can indeed be a bitch!

LOL. I've set it running again on lowest possible stake. Will keep you informed ;)
Post Reply

Return to “Tennis Trading”