Trading What I see !?

[MAGTRADEUK]

So did quite a bit of backtesting last night and came up with the following, small %turnover, but reasonable throughput and happy as I can be it is not too fitted. Nice to be dealing with a +ev backtest for a change! Anyway, going to forward test with at least some confidence now and see how I do:
Profit = 3.99
Wins = 75
Avg Win = 0.26
Losses = 44
Avg Loss = -0.36
S/Rate = 63%
Total tr = 119
% turnOv = 0.23%

How's your progress, Goat?

hi, so I am now running with my new backtested strategy, and seeing how it progresses, I will give an update end of week. It has a %turnover of about 0.4%, which is not great, but I don't think I can squeeze much more out of it. Backtesting also shows quite a variance over 2 weeks, some good days, same poor days, so for example yesterday it was poor, but my testing tells me that's just the variance... Interesting, it seems to perform wonderfully on Saturdays according to the backtest, by quite a considerable margin.

Interesting, so how do your numbers/ ratios change if you only run it on a saturday?

[goat68]

How's your progress, Goat?

hi, so I am now running with my new backtested strategy, and seeing how it progresses, I will give an update end of week. It has a %turnover of about 0.4%, which is not great, but I don't think I can squeeze much more out of it. Backtesting also shows quite a variance over 2 weeks, some good days, same poor days, so for example yesterday it was poor, but my testing tells me that's just the variance... Interesting, it seems to perform wonderfully on Saturdays according to the backtest, by quite a considerable margin.

Interesting, so how do your numbers/ ratios change if you only run it on a saturday?

So with the last two adjustments to my bot the performing days are now noticeably the backend of the week thurs,fri,sat. However, from the performance so far this week i'm worrying i've back fitted, this week so far has been poor compared to the previous 2weeks, but the next 3 days are key I guess, and if that pattern continues, maybe only targetting the end of the week maybe an option. In reality I need far more weeks of data to better analyse...
Here is the complete backtest to-date, the last 3 days this week equal live figures for this bot:
06/03/2021 : Sat : 3.26
07/03/2021 : Sun : 1.01
08/03/2021 : Mon : 1.18
09/03/2021 : Tue : 0.98
10/03/2021 : Wed : 1.83
11/03/2021 : Thr : 4.64
12/03/2021 : Fri : 2.06
13/03/2021 : Sat : 0.20
14/03/2021 : Sun : 0.28
17/03/2021 : Wed : 0.26
18/03/2021 : Thr : 1.34
19/03/2021 : Fri : 4.72
20/03/2021 : Sat : 0.38
21/03/2021 : Sun : -1.58
22/03/2021 : Mon : -2.33 "live"
23/03/2021 : Tue : 0.27
24/03/2021 : Wed : -2.53

==> Worst day = -2.53
==> Best day = 4.72
Sun : -0.30
Mon : -1.15
Tue : 1.25
Wed : -0.44
Thr : 5.98
Fri : 6.77
Sat : 3.84

Profit = 15.96
Wins = 123
Avg Win = 0.48
Losses = 84
Avg Loss = -0.52
S/Rate = 59%
Total trades = 207 from 578 analysed markets
% turnOv = 0.35%

[goat68]

Rather disappointingly the losing streak continues, here's hoping i've not back-fitted this one!
06/03/2021 : Sat : 3.26
07/03/2021 : Sun : 1.01
08/03/2021 : Mon : 1.18
09/03/2021 : Tue : 0.98
10/03/2021 : Wed : 1.83
11/03/2021 : Thr : 4.64
12/03/2021 : Fri : 2.06
13/03/2021 : Sat : 0.20
14/03/2021 : Sun : 0.28
17/03/2021 : Wed : 0.26
18/03/2021 : Thr : 1.34
19/03/2021 : Fri : 4.72
20/03/2021 : Sat : 0.38
21/03/2021 : Sun : -1.58
22/03/2021 : Mon : -2.33 "live from here"
23/03/2021 : Tue : 0.27
24/03/2021 : Wed : -2.53
25/03/2021 : Thr : -1.45

==> Worst day = -2.53
==> Best day = 4.72
Sun : -0.30
Mon : -1.15
Tue : 1.25
Wed : -0.44
Thr : 4.53
Fri : 6.77
Sat : 3.84

Profit = 14.51
Wins = 129
Avg Win = 0.47
Losses = 88
Avg Loss = -0.52
S/Rate = 59%
Total trades = 217 from 607 analysed markets
% turnOv = 0.30%

[rik]

time to up the stakes?

[goat68]

time to up the stakes?

Hmmm

[goat68]

time to up the stakes?

You would think the backtest showing 13 days in a row in profit would be significant, so here's hoping losing 4 of the last 5 days is insignificant!!

[Emmson]

You are very analytical goat which should serve you well, I wish I had some of that myself. Different ways to skin a cat

[ShaunWhite]

time to up the stakes?

You would think the backtest showing 13 days in a row in profit would be significant, so here's hoping losing 4 of the last 5 days is insignificant!!

If you run a Monte Carlo sim using your stats to date you'll see what a wide variety of results the last 200 markets could have given you.

Also as you know your strike rate and averages you can work out the probability of having particular losing/winning days or sequences. Then by setting yourself degrees of confidence your strategy is working or failing (say an 80 or 90% confidence), you can make a decision to scale up or down each day/week. It's data based automation goat, use the same rigour at every stage of the process rather than analysis by data and then implementation/execution based on guesswork or feelings.

[rik]

How to run a Monte Carlo simulation, you have a tool for that?

[firlandsfarm]

How to run a Monte Carlo simulation, you have a tool for that?

rik, have a read here and here. As MCS's are specific to what you want to test there isn't really a one tool fits all but they are not difficult to set up, have a look here (take care not to fall asleep! ) and Google "monte carlo Simulation Excel" for other examples.

Edit, just found an MIT video on MCS.

[goat68]

time to up the stakes?

You would think the backtest showing 13 days in a row in profit would be significant, so here's hoping losing 4 of the last 5 days is insignificant!!

If you run a Monte Carlo sim using your stats to date you'll see what a wide variety of results the last 200 markets could have given you.

Also as you know your strike rate and averages you can work out the probability of having particular losing/winning days or sequences. Then by setting yourself degrees of confidence your strategy is working or failing (say an 80 or 90% confidence), you can make a decision to scale up or down each day/week. It's data based automation goat, use the same rigour at every stage of the process rather than analysis by data and then implementation/execution based on guesswork or feelings.

Thanks Shaun,
So as articles point out of Monte Carlo sim, https://www.pinnacle.com/en/betting-art ... WJHUW5LUZ5 "assuming, of course, my forecasting model was working as it should". These are based on "betting models", which then by using MC you can evaluate if you're current live trading is just "unlucky" or not and to what degree...
In the 2nd article, it uses the model:
"Rosberg, Hamilton and Ricciardo have won 4, 7 and 3 of the 14 Grand Prix held to date in 2014. We can therefore use the ratio of strength 4:7:3:1 for the drivers Rosberg:Hamilton:Ricciardo:Others."
It then bases the MC forecasting on that, so all the results are based on this model, which is obviously flawed because Rosberg & Hamilton were on the same team, so the 4 and 7 weightings will not be independent variables, especially going into the final races...but then that's saying maybe I could have come up with a better model!

I don't have such a "model", i've no idea what is "fair value" from which I can model an expected return...be that good or bad!!
Your point about strike rate and averages is also interesting and something I am looking at, however you can only go forward and decide to scale up or down if you have a "model" based on the data to reasonably go by, which is my intention. However, at the moment, and this is probably glaring obvious from my stats above, i've not got enough data to go by yet to have a good enough informed model.
However, having said all that... this is my current set of observations:
- My judgement at the moment is the results this week are most likely within the realm of the lower quarter of the bell curve (-1 SD), so about a probability of 1 in 6, but this is just "judgement" given I can't really define "fair value"
- My data now covers about 3 weeks, however what's key is if you look at what those 3 weeks are:
week 1: pre-Cheltenham
week 2: Cheltenham
week 3: post-Cheltenham
As you can see "Cheltenham" bang in the middle, and I know this is postulating, but I very much suspect the "betting" behaviour for each of those weeks is different. This I suspect is probably at the moment the biggest influence on my "results".

[ShaunWhite]

Edit, just found an MIT video on MCS.

Thx for fielding that.

[ShaunWhite]

I don't have such a "model", i've no idea what is "fair value" from which I can model an expected return...be that good or bad!!

There's elements of a true MCS that won't apply but you'd been mentioning your return on investment, that's the addition value your model/strategy/algo appears to be finding.

In its simplest form you can setup a spreadsheet that has 100,000 lines of a 50/50 random coin flip, and assume a return of Eg 50.1 for a win and - 49.9 for a loss etc.... Then bash away at the F9 recalc key to see how returns can vary. Just for the purpose of illustration rather than giving to any hard facts of course.... Or instead of a 50/50 flip, use your strike rate and instead of 50.1/-49.9 use the return you've found so far.

Don't ask me for any more info, my maths never got beyond an 'U' ngraded A-level in 1982.

[goat68]

So just checked in on bot's progress today, currently -£2.37, 6 losses out of 7 ! Hmmmmm, will see how it pans out by end of day.
It is starting to look like I fitted a strategy to the "Cheltenham period"... anyway will keep running it, could be just a drawdown....
I do know from my stats it would only take a couple of good wins to make today +ve.

[ShaunWhite]

So just checked in on bot's progress today, currently -£2.37, 6 losses out of 7 ! Hmmmmm, will see how it pans out by end of day.
It is starting to look like I fitted a strategy to the "Cheltenham period"... anyway will keep running it, could be just a drawdown....
I do know from my stats it would only take a couple of good wins to make today +ve.

6 out of 7 is no big deal. I think my record on a strategy with a similar strike rate to yours is 13 straight losers and 22 straight winners. Daily results really aren't worth looking at. Even 6 in a row on a 50/50 is only 64/1 and when you trade a lot those come up pretty often.

Re fitting, that's why you need to do your analysis in one dataset and testing in another. Those sets could be different days or better still something like alternate races. You can't just try and find a strategy in your whole set because there's always going to be one in there somewhere. In zero sum data all you have to do is exclude a few losers and it looks like you found something.