Future of Day Trading = deep neural networks
Finally (check my blog post from 2018) I have managed to finish one of my eternity project to some form, where I can start to make trades based on AI rules automatically. Actually a deep learning neural network has been taught to do day trading and finally results seem to me more than great. That has an implicit rule that stock market movements aren't random walks inside a trading session! (if it would be a random walk, there wouldn't be any possibility to find out any working model but let see how well this model start to generate alpha, I know lot of those results will be vanished with technology issues, bad feed and other common mistakes) Reason that it can generate alpha is that statistically market movements don't follow standard deviation but have long tails and so on.. that is already proved and and this project will show how to utilize it.
I'll start to trade E-mini (NQ, YM, RTY tested, NQ is the first one in the list) futures publicly beginning of May 2022 so this is announced before I have even started to trade that system, arrogant!? Maybe but check the result later!
Showcase:
System will be published in C2 and most probably somewhere else which tracks real time trades. Wait a direct link starting of beginning of May. Update 2022/05/02: so here it is AI PV Mini Futu (First quick MNQ trade was just to test NinjaTrader link)
There will be a limit to followers if this goes too viral. The main goal is to fill my own account, shame on me.
Probably it is better not to show those backtest out of data results as nobody wouldn't believe those, even I'm a little bit suspicious if there is a data leak somewhere ... But this will be seen and back to the drawing table or retire then.
What should be expected:
- Day trading
- Lot of small negative trades,
- Larger winning trades
- Momentum trades with early entry
- Winning trades percentage less than 40 %
- No fixed target and stop losses (You can add some higher fixed ones for safety if technology issues)
- No overnight (over session) positions, lower margins needed
- No martingale, only 1 contract of each futures traded (Own note for future: lisää toinen higher profit facktor myöhäisempään aikaan eli n. 2 pf ensteks ja sit n 3 pf (tai toisinpäin), kun ensteks kerännyt tuottoa ja kullekin instrumentille omat systeemit sekä laskenta puhtaasti ninjaan nopeesti)
- Couple of trades per day
Technology behind:
NinjaTrader is used to handle the feed, actual trading and transfer data between backend python program (tensorflow keras Conv2DLSTM model.. The current model has about 3400 parameters, 9 features used.. I know, you say overfitted but pure out of data tests done) which do prediction based on data feed which is preprocessed by NinjaTrader 8. Initially NinjaTrader is used to duplicate trades to C2 too. (some delays coming from the start of python and load of tensorflow + model, no idea how this will affect the final results, at least some slippage need to be added) Lot of those adaptive indicators and tools used like PVAMA, PVPMAdaVTR so those might have a lot of effect when parameters optimized properly.
I have to show a picture from my whiteboard, which has a very old draft from my early deep learning models at left down corner but added that Conv2LSTM there last autumn and that is the key part now but more important is what to predict!, not revealed, sorry). Quite a mesh, to be honest. Five years of heavy and sometimes much more occasional development behind. Countless amount of coffee.
Other:
No, this is not for resale, too complicated to teach and use and lot of basic knowledge needed with deep learning before utilizing although all the major parts has been there since 2019 (some crab, some brilliant at least idea wise). Not reasonable packed together, not even now so don't do anything else than look at it ..
Reason that this has been delayed so much is that I have had a lot of customer projects at fiverr making hundreds of conventional trading robots and indicators, be free to find out PVoodoo from there. I'm proud of the customer feedback there, but sorry my customers, maybe I have to postpone your projects again..
Bookkeeping (Notes for me until everything cleared):
2022-05-02: Wrong bar type at NQ initial setup -> overtrade fast movements (but similar drawdown can happen at correct one, too. Just FYI)
2022-05-03: Lot of those small negative ones, I can't see positives...
2022-05-(04-05): It is heavily overtrading NQ, still (so some difference between trading machine and development test machine) (NQ, to hold as long reason found) Reason found: wrong h5 file, shit, better naming needed, start again. Better last entry & exit times set (for better liquidity and for no spread)
2022-05-11: Platform transmit off, shit, (during the most well behaved nice long swing) Corrective action: windows updates paused for right possible update date.
2022-05-31: It seems that it look like a major catastrophe here (typical one, just opposite ones would have performed much better) But technology tested and it seems to work fine. But some hold and checkup with settings now!
No comments:
Post a Comment