Polarity ATI July 16 Power Hour
Вставка
- Опубліковано 5 вер 2024
- Auto Arm regression model July 16th for some ML time on late afternoon institutional order matching.
Was not as much action as I anticipated but ML time is good.
If you like the channel content, please subscribe and give it a like.
Thanks for stopping by!
Brett,
Thanks for sharing all of this information. It's amazing how the slightest adjustments make all the difference in the world.
they are sensitive.
Awesome ML brett, great work!!
thanks
Any reason why the Bar graph dont appear right after training the model telling me the results on a separate tab like the previous version? And how come mltrain or any data dont appear in my pati folder at all?
it could be a lot of things. without seeing exactly how you built the model, it's hard to diagnose the issue.
hi Brett, i'm looking for models with low RMSE and high R-squared, but i would like an example in number. for example rmse: 8.63 and r2:0.02 is good? thank you
depends on what kind of model you are building. If building a winrate model, I wouldn't worry about rmse / r2.
If building a number of trades model, then rmse / r2 are much more important. Here I would shoot for rmse under 15 and r2 over 40. This is easier accomplished by appending a train the model. follow these steps:
Append directions for building a model Building model with more than 1 chart data set.
Load a 3 / 5 minute (or whatever type data series you want) chart with 20-30 days of data. Make sure your indicators you want to use are loaded on to the chart.
enable your model in train the model mode with Overwrite ON
run 'Train the Model'
get rmse and r2 make a note of the result
***with model still active***, open data series on your chart and switch to tick or range chart you plan on trading the models with.
let it load to train the model
disable model, turn off overwrite. re-enable model
run 'Train the Model' again.
should be much improved rmse and r2 - results will depend on how well you choose your indicator set.
rinse / repeat for regression model
Do you think you will try and emulate Connor's Hyper-scalping strategy?
I am working on it. New series starting today on ES Pre-pack
Hi Brett, what Sampling Interval & Trades Window have you found that it is good for NQ and ES
NQ is fast paced market. I use 15/20 second and 200/300 window.
ES is slower. 30 second and 300/600 window depending on what type model one is building
Nice win rate on this session.
If I want a model for trading a certain time of day, would it still be best to incremental learn on the full RTH session to provide more learning opportunities or be specific with the timings? Cheers
depends. if your trading window is around the same time each day to fit your personal life schedule, then I would just ML on that timeframe. If you are around the markets all day, then RTH ML is good.
Brett i have watched all the videos you posted and made notes on most of the info provided by you. i have few questions
1. when i create a new model in ninjatrader what are the files which needs to be created in PATIResources folder automatically.
2. how can i create files in Patiresources folder if the files are not being created automatically?
3. when i train in trend mode the model is getting trained but if i try to train switching it to regression it is not getting trained. and in the output it shows as ML data and/features files were not found. Please ensure tick replay is ON (this is the latest MLPRELEASE version)
4. when you say we have to take good data before it goes to catboost, do you mean form the connorML Train files? if yes then will there be a new file for each models created or the regular connor ML train file?
The files in that folder are automated. There is step by step tutorial in the included PDF file. The only thing you do is make sure PATIResource folder is available and empty when you start.
throw a PATILapp.exe in there. enable that. then add new models with Polarity Models builder interface. one for trend and other for regression.
load indicators you are using to chart.
add a strategy to your chart. adjust settings to how you want to train the model.
down about middle of settings 'Active for training' is the trend model. make sure that model is in there. regression model goes underneath in next window.
enable model. MLTrain file will auto populate in your PATIResource folder.
hit 'train the model' on your chart trader. let it complete.
disable the strategy. scroll down and put the regression model where the trend model was located 'Active for Training' can leave regression slot below it as well.
right click chart, reload ninjascript. this reloads the indicators for regression model.
enable strategy again. you will see another MLTrain show up in PATIResource folder with different serial number.
hit 'train the model' on chart trader. let this complete. you can watch progress on open PATILapp.exe window.
when that completes, you should now have (2) .cbm files in your PATIResource folder. One is trend, other is regression.
disable strategy, go back to 'Active for training' window and put trend model back there. leave regression below it alone.
turn off Training mode and overwrite
turn on optimize settings and incremental learning.
you should be set from there.
all this info is in the pdf file from FA
@@BrettR-Trades i tried doing it but the heuristic parameter optimizing button doesnt appear when i try training the models i create and it shows as .csv file (model for regression) is deleted and not found. but when i train for the models in prepackage, the optimization button appears.
when i asked Pete he asked me for logs and when i sent him all the files and errors from output, he asked me to delete and create again though i have done it many times.
thank a lot for your the instructions and information, i really appreciate it and currently i have got RMSE 6 and R2 0.7 however, there are lot of thing i.m not familiar with
It takes time. lots of time. I knew nothing about Polarity when I purchased other than watching Connor and Ron rake big pots on their YT posts. It's important to learn what the files do, what they are for, what the settings do and how to adjust, etc. it's a lot to digest.
Hi Brett, do we need to train a new model for every contract rollover?
if a model was built with March contract data only, and only ML optimizing since, then machine learning on June / Sept contract is fine on that March build model.
problem arises when building a new model right after contract roll and mixing new contract data with "back-merged" re-calc'd previous contract data because of the large spread between price on the 2 contracts. Just build 'train the model' a model with one contract data, then can keep around as long as you want with machine learning afterwards. just the way I choose to do it. others may have other opinions.
Can you please make a tutorial on training the model?
tutorials are time consuming. hopefully the FA team has one coming soon.
How do you go about deciding when to do Regression vs Trend? Thanks.
each model needs experience in all market conditions when machine learning / optimization. if a market is trend ripping, and auto arm is in regression mode, we want that model experienced enough to wait for a proper volume imbalance / ratio ie pullback to run over trail stops to take an entry and run some stops with it. trend needs experience being patient in rotational day markets. watch some of Ron Klemps videos. he will set a trend model in a tight range rotational market and it churns out winners in small windows. His models have a lot of ML hours on them to learn that
@@BrettR-Trades that's a good point. I always wonder that when watching his model. If your model takes bad trades, do you go and filter those results in the ML file ?
If you are using the latest Polarity update released last week of June, you do not need to filter the Incremental file. It collects much better filtered data now. In my opinion, that was the biggest improvement in the last Polarity update.
Brett, When creating the inc master, do you do small 2-3 hour sessions over a few weeks of data, then combine, or do you load the whole time frame and run it creating a very large single incremental.
Have you noticed any effects of speeding up the playback on the incremental learning? I vaguely remember Ron saying x8 was the fastest he would go. I will have to try and revisit his videos.
you can combine several sessions into one then upload. My minimum is usually 4-5. I try and keep that catboost forest from getting to thick
starting out I would let every session upload. I had cbm files that were huge (well over a million. some 2 million), the larger they got, the worse the winrate became. that's when I started researching catboost. that was a no no the way I was doing it.