ProRealAlgos? Real or fake
- This topic has 226 replies, 29 voices, and was last updated 11 months ago by
.
-
What’s going on here… Maybe some live data should just be shared here to refute any allegations? That would be proof enough that it works… or not. It has been a difficult time for many algos on NASDAQ and SP500 since July. It goes sideways with a downward trend. So a difficult environment for algos that were developed in smaller TFs and tend to be over-optimized and are now collapsing. By the way… 2 of the 3 shared algos in their library are crashing right now. But… that happens with pretty much every algo right now in M1 or M5. I can say from my own experience that there are very few who continue to make good profits. But now I know what I have to pay attention to. So share some live data as proof…. otherwise some people here never let up. Whether good or bad remains to be seen.
1 user thanked author for this post.
1 million bars is available under PRT Premium. I use it to test with 6 months 1minute data, to next discover that the data beyond that (towards the past) in 100% of cases is overfitted. I know that it happens – I know I need to avoid that, but it happens anyway, always. That’s why my question to Carl (which is rhetoric).
I sat down to analyse what all to avoid (which in the end is all obvious) and recently I developed systems which work at the second level – pure technical systems. Message here : with 3 seconds and 1M available bars, I use 3 weeks to test and tune the system and I have ~6 weeks more for sanity checking the overfitting.
More should not be necessary because I now make all independent of markets. The tip I already have by now : don’t expect many trades (like maybe 1-2 per day) and don’t expect much profit (in points) as well.Note : We can’t test each other’s overfitting, like I could let loose 1M bars on your System, unless it is free of Market Place royalties. But you can do that yourself :
Just run the past over it for as much as you have (like 200K bars) and observe the “bend” where DD starts. Now bet with yourself that that was the date the particular system was ready and put to the Market Place. So what you can see from the past – where the system started to be developed (like in your case today 200K bars ago) will happen just the same for the future – you put your system Live 2 months ago, then backtesting it today will show the peak at exactly 2 months ago. OK, you knew that. But it is telling about 3rd party Systems. You could show us a few, for mutual fun. But if you can’t … hey … then Carl is right. But if not, then Carl, no stories / excuses about coincidental markets please. Remember what I told : this is all to be expected. With you – with everybody/every instance.🙂
1 user thanked author for this post.
Thank you Allan. Let’s agree to disagree.
Coincatcha, yes you can get access to a million bars with ProRealTime premium as Peter said. With that you should get 9months of data on the 1 minute timeframe, compared to 2 months of data with the basic version. I’m sending you the link here: https://www.prorealtime.com/en/ProRealTime-Premium.
I agree, I think us sharing our approach to avoid overfitting and our development process in general could help others. We’ve done some videos on the topic of algo development and more is coming.
I will try to write a more complete thread on the topic later. Any way, on this topic you can watch a video here where we are covering the essentials of optimization https://youtu.be/Pu1P7o3syHM?si=XomwIjFqlfnw-x4EPeter is on to something here on the topic of evaluating algos that I agree with. The best way to evaluate and check if a system is overfitted is to first verify which date that EXACT VERSION of that EXACT ALGO was released. Some algo developers are continously releasing “new versions” of their algos and then attribute/claim historic results of another version to this new version saying that it’s been live for years. A new version should be considered a new system and should be evaluated accordingly. So, find out when the latest “version” was released, save it for a couple of months and then run a backtest of it and evaluate the results from the release date until today. THAT’S THE ONLY RESULTS THAT MATTER. This approach will save you alot of €€€ from not running overfitted systems.
1 user thanked author for this post.
Personally, I have written many, many algos, sometimes x different versions of the same procedure. When I look through my library from time to time and pull out systems again after x months, I see how they have developed without being touched. This is how I figured out useful things for TFs in 1, 5, 15 minutes. What could matter. For example, it seems to make sense to consider a high TF from H4 for a trend. Or fixed sizes like Dhigh/Dlow/Dclose, or a weekly close. Things that are always there and cannot be optimized. If you build systems around it, you have a high chance of getting something robust. These are my findings from the last few months.
2 users thanked author for this post.
It is often very beneficial to allow a maximum of one position per day. Especially in sideways phases, a trend system loses because it tries to make a profit in every direction but is caught by the SL. Back and forth makes pockets empty.
1 user thanked author for this post.
Find exclusive trading pro-tools on 
