Hi all, I have been using an automated bot for a few weeks, and it has been profitable with only small stakes. However, over the last week, I have noticed the profits getting smaller and have even started losing money.
I guess this is pretty standard as markets adapt?
Markets Adapting to My Bot?
-
ShaunWhite
- Posts: 10494
- Joined: Sat Sep 03, 2016 3:42 am
I'd upgrade Peter's 'unlikely' to 'no', if it's just small stakes for a few weeks. And as well as being a good strategy that's having a dip, it might just be that it wasn't an edge in the first place and gains were variance.
And worth remembering there's dozens of people starting and stopping new automation everyday so markets evolve, and strategies can fade, reverse or die periodically, or terminally. But they can also be long lived, I've got one that's been struggling for a few months recently after being like clockwork for 6 yrs.
A pl chart would make it easier to judge the effect you're seeing.
And worth remembering there's dozens of people starting and stopping new automation everyday so markets evolve, and strategies can fade, reverse or die periodically, or terminally. But they can also be long lived, I've got one that's been struggling for a few months recently after being like clockwork for 6 yrs.
A pl chart would make it easier to judge the effect you're seeing.
-
ShaunWhite
- Posts: 10494
- Joined: Sat Sep 03, 2016 3:42 am
Tough question with so many considerations. If it's not just pissing money then drop down to the smallest stakes you can and consider the small losses as the cost of gaining information.
How many bets (not markets) a day is it doing? If its a couple of hundred then you might get good info in a week.
But I can't recommend anything better, unfortunately this is the problem when you don't have a way to re-run variations on the same period, and have to wait so long for evidence. So my solution about 8yrs ago was to bite the bullet and switch to a more specialised bespoke automation setup. It took the best part of a year to setup all the functionality I needed, but manual traders expect to take a year to get going so it didn't seem unreasonable and improved the chance of success significantly. It was a gamble but that's the game.
The guys who use BA automation will probably know more about how you solve your problem, because I couldn't.
-
jamesedwards
- Posts: 4117
- Joined: Wed Nov 21, 2018 6:16 pm
It is possible if a bot works in a way that can be easily spotted and manipulated.ShaunWhite wrote: ↑Tue Apr 29, 2025 12:45 pmI'd upgrade Peter's 'unlikely' to 'no', if it's just small stakes for a few weeks. And as well as being a good strategy that's having a dip, it might just be that it wasn't an edge in the first place and gains were variance.
And worth remembering there's dozens of people starting and stopping new automation everyday so markets evolve, and strategies can fade, reverse or die periodically, or terminally. But they can also be long lived, I've got one that's been struggling for a few months recently after being like clockwork for 6 yrs.
A pl chart would make it easier to judge the effect you're seeing.
eg I once wrote a greyhound bot that tracked and offered best back and lay prices in early forming markets. It was making very good money until someone realised they could manipulate my prices and then snaffle the value. Profit suddenly took a huge dive and I wondered why, then saw it happening in real time.
-
ShaunWhite
- Posts: 10494
- Joined: Sat Sep 03, 2016 3:42 am
Similar, but I don't think it was responding to me necessarily, just someone launching a gap filler and we just stepped over each other until the spread was gone and the cycle repeated. There's dozens of people doing random things and everything has an effect. Chaos theory, one new bot causes ripples and we'll never really know what's caused what.jamesedwards wrote: ↑Tue Apr 29, 2025 2:35 pmeg I once wrote a greyhound bot that tracked and offered best back and lay prices in early forming markets. It was making very good money until someone realised they could manipulate my prices and then snaffle the value. Profit suddenly took a huge dive and I wondered why, then saw it happening in real time.
High variance bets (e.g., betting on long odds) require a larger sample size to overcome short-term fluctuations compared to low variance bets.
If your strategy has a very small edge (e.g., 1-2% +EV), 1300 bets are likely insufficient to distinguish that edge from random noise (variance). Variance can still heavily influence the results over this sample size. You need a much larger sample size (likely tens of thousands of bets) to be more certain that your results are due to skill.
If your strategy has a very small edge (e.g., 1-2% +EV), 1300 bets are likely insufficient to distinguish that edge from random noise (variance). Variance can still heavily influence the results over this sample size. You need a much larger sample size (likely tens of thousands of bets) to be more certain that your results are due to skill.
-
jamesedwards
- Posts: 4117
- Joined: Wed Nov 21, 2018 6:16 pm
I would say no. But 1300 is a great start and definitely promising.
For a basic statistical view using regression analysis check your R². Plot your results on an excel graph, right click, add trendline, check the 'display r-squared value on chart' box.
The closer your R² gets to 1, the greater the likely significance of the data. Anything above 0.7 shows some decent statistical relevance.
You do not have the required permissions to view the files attached to this post.
-
ShaunWhite
- Posts: 10494
- Joined: Sat Sep 03, 2016 3:42 am
I'm going to sound like a broken record now but ....
1. Download your bet history.
2. Add a running total column.
3. Make a line chart of it.
In normal date order it's going to be a rough line with big up and down spikes as it's per bet not per market.
4. Add a trendline, use a moving average over enough samples to give you something smoother.
5. Now sort your bet history by col C - BidType.
Are the backs and lays both upward lines on the chart?
....or is it like this where the backs make money and the lays give it back? 6.Now sort your bets by col K - Avg odds matched
Is the chart good across the whole price range?
You get the idea.
The bet history also has the event time and the bet time....from those you (excel) can workout what the countdown timer was when you placed the bet. Sort by that and see if you're good early or late. With some work you could add the BSP, and with that you can work out the EV on each bet which is much better than the cash which can be misleading. Anything you tie back to a market or individual bet can be useful.
With a couple of excel templates and macros you've got a half decent bet analyser that's easy and quick to use. Ta dah!
The bottom line is the same however you sort it, regular time sequence tells you overall performance over time but nothing else useful. Market totals don't mean much either if you automate and check your PL less often so you soon start to see it as a stream of bets not 'trading markets'. Analysing bets from every angle and thinking of new angles is pretty much what we do all day while the other guys are having fun watching sport
New automators might not realise what you do all day while a bot is running for a week.
You do not have the required permissions to view the files attached to this post.
-
ShaunWhite
- Posts: 10494
- Joined: Sat Sep 03, 2016 3:42 am
i just remembered this post ....
You can calculate them all in Excel with some AI help
How much of the variance in the data is explained by the model
R² Value Interpretation
≥ 0.90 Very strong relationship —the model explains almost all variance
0.75 – 0.90 Strong relationship —good explanatory power
0.50 – 0.75 Moderate relationship —some noise, but still useful
0.25 – 0.50 Weak relationship —model only explains part of the variance
≤ 0.25 Very weak or no relationship —mostly random noise
How far the observed mean is from zero, scaled by variance and sample size.
t-Statistic Interpretation
≥ 10 Extremely strong evidence —almost certain the effect is real
6 – 10 Very strong evidence —high confidence that this isn’t noise
3 – 6 Strong evidence —significant, but not bulletproof
2 – 3 Moderate evidence —likely real, but worth further validation
1 – 2 Weak evidence —could be noise or require more data
≤ 1 No real evidence —likely just random fluctuations
The probability of seeing this result if the true effect was zero.
p-value Interpretation
≤ 0.001 Extremely strong evidence against randomness —highly significant
0.001 – 0.01 Very strong evidence —unlikely due to chance
0.01 – 0.05 Strong evidence —statistically significant
0.05 – 0.10 Some evidence, but weaker —possibly real, but less convincing
0.10 – 0.20 Weak evidence —could be noise, needs more data
≥ 0.20 No real evidence of a meaningful effect —likely random
You can calculate them all in Excel with some AI help
How much of the variance in the data is explained by the model
R² Value Interpretation
≥ 0.90 Very strong relationship —the model explains almost all variance
0.75 – 0.90 Strong relationship —good explanatory power
0.50 – 0.75 Moderate relationship —some noise, but still useful
0.25 – 0.50 Weak relationship —model only explains part of the variance
≤ 0.25 Very weak or no relationship —mostly random noise
How far the observed mean is from zero, scaled by variance and sample size.
t-Statistic Interpretation
≥ 10 Extremely strong evidence —almost certain the effect is real
6 – 10 Very strong evidence —high confidence that this isn’t noise
3 – 6 Strong evidence —significant, but not bulletproof
2 – 3 Moderate evidence —likely real, but worth further validation
1 – 2 Weak evidence —could be noise or require more data
≤ 1 No real evidence —likely just random fluctuations
The probability of seeing this result if the true effect was zero.
p-value Interpretation
≤ 0.001 Extremely strong evidence against randomness —highly significant
0.001 – 0.01 Very strong evidence —unlikely due to chance
0.01 – 0.05 Strong evidence —statistically significant
0.05 – 0.10 Some evidence, but weaker —possibly real, but less convincing
0.10 – 0.20 Weak evidence —could be noise, needs more data
≥ 0.20 No real evidence of a meaningful effect —likely random