gmg733
788 posts
msg #135706
- Ignore gmg733
modified |
5/1/2017 4:50:22 PM
NN was next on my list of trading research after Nadex trading. I started looking into Neuroshell and Wave59. Didn't know if you had an opinion on these or if you had a recommendation? I know the Wave59 guy is out of CO where Stratasearch is HQ'ed.
Sounds like some great research. Great job as always!
Thanks.
|
Kevin_in_GA
4,599 posts
msg #135707
- Ignore Kevin_in_GA |
5/1/2017 7:46:22 PM
Neuromaster 3.3 Pro - while not having used this for very long yet, I am pleased with how quickly you can get up to speed and it already paid for itself with this mornings trades. The developer/programmer is fairly responsive but not as good as Pete at SS. There does not seem to be a forum for users, so that is a bit of a minus.
|
gmg733
788 posts
msg #135709
- Ignore gmg733
modified |
5/1/2017 11:13:28 PM
Thanks. Yeah Pete is great. Nice trading btw.
I was thinking too, is it safe to assume the 10 systems have a lot of the same entries and exits? The results are what they are and give some diversity. I was just wondering.
|
Kevin_in_GA
4,599 posts
msg #135712
- Ignore Kevin_in_GA |
5/2/2017 7:55:40 AM
Well, only two of ten triggered for yesterday. Before that two others were in play, with one closing on 5/1 and the other closing this morning. The others are on the sidelines right now. To me that seems like some diversity of entries and exits. I have all of the entries and exits for each system, so I could check.
|
mahkoh
1,065 posts
msg #135721
- Ignore mahkoh
modified |
5/2/2017 12:59:22 PM
Curious to find out what you think of it.
The website looks a bit like a pennystock newsletter, but who cares as long as it works?
|
gmg733
788 posts
msg #135725
- Ignore gmg733 |
5/2/2017 7:25:33 PM
Thanks Kevin.
|
nibor100
1,010 posts
msg #135732
- Ignore nibor100 |
5/3/2017 5:04:12 PM
@Kevin,
Regarding your initial 10 NN Models, on examination of the backtest data you provided it seemed to me that there might be more variability between the input and exit criteria of the different models than it seemed at first glance.
After all, picking the best performers out of a larger group is probably going to force some correlation into the mix , as those models are each more likely to have greater numbers of the best trades during the test period.
1. In the 10 models, the results show that number of trades varies from 25 to 34 for the different models, which results in entry criteria that create several instances of 10% or more trades, between models, up to a max of 36% more trade entries.
2. Using the avg days in trade for winners and losers times the # of winning and losing trades indicates that during the test period the modes ranged from 256 days to 325 days of XIV ownership. Curiously, the model with only 25 total trades and no losing trades had the 325 day market presence.
3. Since there were at most 2 losing trades in any of the models, the avg # bars in losers seems to show quite a difference in losing trade entries and exits. Two of the 1 loss models have 13 day losing trade durations, while several of the 2 losing trade models have an avg of 4 or less days per losing trade.
4. I was going to look at some other stuff related to loss and win size, and consecutive winner placement but I see that you have moved on to a new set of models.....
Note: Your one and only live trade seemed to be out of place when compared to the model results by having an avg trade duration of 1 day and and lower avg winning trade of $623. ( I used starting portfolio values of 10k,20k, and 30k for the avg to be a little more fair.
Ed S.
|
Kevin_in_GA
4,599 posts
msg #135733
- Ignore Kevin_in_GA |
5/3/2017 7:10:11 PM
The composite stats on those filters were a bit misleading - the difference in total number of trades arose from several systems moving in and out of trade periods held by the others (e.g., two short trades over the same period other systems held one single longer duration trade). There were very few instances where one system entered and exited a trade without any of the others also entering at nearly or exactly the same time. And average trade stats can be skewed by one long trade hold that made more money than others.
I think the new set of filters, where the neural network parameter space has been more thoroughly explored, will be more in line with my goal of having multiple low-correlation NN systems all trading profitably (in essence, covering more profitable trade periods than just one system, and allowing me to ease in and out of trades rather than go all in with just one model).
Still way too early to say if this approach works as I hope it will. Only time will tell.
|
graftonian
1,089 posts
msg #135746
- Ignore graftonian |
5/4/2017 2:45:15 PM
Kevin,
Thanx for posting your results of NM trading. I have downloaded the free version, but it is not up and running yet. Waiting for reply from their support staff. Were you able to get it running "out of the box"?
I am getting "downloading failed " and "unknown file type" when updating, re-learning, or creating new models.
Graf
|
Kevin_in_GA
4,599 posts
msg #135747
- Ignore Kevin_in_GA |
5/4/2017 3:06:29 PM
The free trial is not functional in that it only uses data from 5 years ago - it shows you how to do things but does not import any recent data. I decided to make the investment and for me it paid off in the first trade, but I guess like any trading software it depends on how well you understand it and actually trade the signals.
|