You are, it seems where I wanted to be at the close of 2018 -- but I feel like I'm barely out of the starting gate. Many competing pressures, for sure, but I got spoiled by 2014-2017, when I was wickedly productive on options work. And worse, perhaps, is that what I've gotten up/running for 2018 was based on notes from frikken 2012 -- rewarding that the early work was well-done and not-for-naught, but sooooooo frustrating, that actual *progress* comes now in dribs and drabs. YOU, on the other hand appear to be spanking it. WELL DONE, fan27.
I am also curious how you allocate money given that you trade multiple strategies. I would assume it will mess up the backtests as you may miss trades or take them with smaller size.
Thanks! This project has been my full time job since the beginning of July of 2018 and I was working on it on the side two years prior to that. It has been a huge exercise in patience as the majority of the time my trading account was idle. I am encouraged by the initial results but realize I need more account highs under my belt before my approach is fully vetted.
After I run the OrderOptimization runner I have a bunch of strategy candidates with trade candidates. Depending on the max position values I have set at the individual strategy level and strategy group level, trades will be skipped as max positions are reached. Also I have a account percent risk per trade. Let's say it is 2%. If, for example, my account balance is 100,000 (backtesting or live) and I have a strategy that has an ES trade where the stop order is 4 x ATR(14) where the ATR value is 30 points, then the total risk is 50 (ES point value) * 30 * 4. That is $6000 so my risk component would bypass this trade 6000/100000 = 6% which is greater than 2%. So you are correct in that different initial starting balances for the backtest will yield different results.
Currently I am running AlgoTerminal on my laptop. My next task starting this week is to get AlgoTerminal deployed to the cloud and add monitoring, disaster recovery, etc.
Hence the term, "Machine Learning". [!!!] You've got the algos laid out for choice -- contracting markets, expanding markets, quality pools, copper+transports and "MOMO" waves, right? Well, now that second (or third??) layer to choice between 'em. HA! There goes your 2019! {Hee hee heeeeee!} HA! Our messages crossed, but "Knewwww it! Knewwww it!"
Just went through the entire video. Want to commend you for a very clean architecture, so far as that goes. You followed major conventions, making it very easy to understand even from screenshots. That being said, I don't think you're going to make money selling this for other traders like the ones you see on ET. Hire a salesperson and get them to sell it to small quant funds. Start off at 25-50K/pop, get some feedback and some revenue coming in. Give them the white glove treatment. The reason I say this is that your system is obviously not only professionally done, but needs a professional to make good use of it and see the value. So make sure you choose your target market carefully. The other danger I see with the approach you've taken is that it is so easy to overfit. If I understand correctly, you take a random bunch of criteria, fit them to past data, and then try to apply them to the future. The professionals here can correct me if I'm wrong.
Maybe these guys would be interested in collaboration of some sort. https://www.elitetrader.com/et/threads/iexcloud-io.328989/#post-4792664
Good points. My initial thought is to see if I can get a handful of people willing to pay me for 40 hours of consulting which would involve custom programming, training, etc. and they would get a lifetime license for the tech. This would help fund the venture to the next stage. My ultimate target market is not retail though. If that does not happen, I will likely get another software related job and keep using the tech myself. The criteria is not random as it is user defined and the user will know exactly how the features will be combined. The application just automates the process. Of course any form of backtesting uses past data and is in danger of overfitting. My current approach is to test on multiple slices of data and the application supports walk forward testing.
This is really great, almost too good to be true. But I can see now that it's way achievable in the market to make good steady profits.