P123 gives me exactly what I need,

[Jrinne]

All,

I went ahead and paid the $500 dollars for API credits. Plenty of credits for what I needed.

Here is a k-fold validation with an test sample and embargo period of 3 years (pretty aggressive embargo period) comparing recursive feature elimination and no recursive feature elimination. As you can see the results are similar. I have done other analyses.

But I did a k-fold validation on all of the data after this with no recursive feature elimination (also no test and embargo period). Here is the backtest with the resulting 20 stock sim (using the purely ML derived ranking system):

Everyone (including me) can discuss, worry and/or speculate about what my out-of-sample results will be.

But here is my first point: I emailed a friend this morning and mentioned to her there is no need for me to discuss this in the forum anymore.

I do not need to make a feature request to do pretty much whatever I want to. And I might add that I am more likely to get a direct answer from ChatGPT than the present P123 forum regarding machine learning.

I will also note that only recently has there been a DataMiner download with accurate forward returns. Maybe a month or two? One could say I could have gotten here earlier but could I have with incorrect future returns? And one could say I should have if I were a better programmer. But the simple fact is I am not that great of a programmer.

It has been a struggle for me to get to this point. But I am pretty much done.

I would not mind daily updates of overnight data with in the DataMiner download but I can actually do that now with the API if I need it.

While I am done essentially, P123 will be starting just now with machine learing. I will add this to something P123 will need to consider with regard to the forum.

And this. There are a lot of machine learners at P123. Few of them bother to post. I have noticed more than a few that come discussing their neural-nets usually. But very quickly stop posting.

I leave this purely as a question that probably someone else will have to answer. As I said, I am now one of those machine learner that pretty much can do it on my own now.

I mean that purely as a positive thing for P123.I am thankful for the data. If I have any motive at all it would be that P123 survives as a business for as long as I am investing with machine learning

I am not claiming to have any answers on marketing or what to do in the forum today. I think I will stop with unsolicited advice at this point. I do not get paid for my advice and as I said, I have what I need.

I will leave that now to those who get paid to answer those questions going forward. My only interest being that it ultimately does work for P123 so I can keep using their services.

Jim

[Quantonomics]

I ain’t reading all that. I’m happy for you tho, or sorry that happened.

[jlittleton]

I’m also super happy with the downloads. I’m having a few issues with my downloads not lining up, despite using the same universe, but nothing serious.

They have enabled me to try different optimization methods and validation methods really quickly. This not only makes the engineer in me happy it also makes it very easy to test my investment ideas in a robust way!

My initial tests are showing that weight optimization is not giving me out performance when using time series cross validation. While I want to investigate this result more, it is super useful to be able to research as it tells me I should focus my efforts elsewhere such as factor discovery. Maybe this next month I can do a more scientific study of optimization and cross validation and share my results on the forum.

Anyway, Jim I hope you will continue to share some of your findings, if you are comfortable with it, as your advice has really helped me out so far.

[Jrinne]

Jonpaul,

I am happy to share my machine learning experience with you. I enjoy sharing ideas with such a capable machine learner.

You and Walter seem to be the only 2 interested in the subject. I wonder if it is not too much trouble for you (and of course Walter any anyone else really) to email me with questions.

It just is better with my work situation. I tend to have free time doing some Python on weekend mornings and would likely get back with you immediately then.

Cool what you are doing. The only comment I would have is that like you, I have it all planned ahead on what I will do and then test something that sends me in a different direction.

The above model for example was the second I tried. The first worked well. The one above was more of a “I will do it just because I can in about 2 minutes of computer time.” And here I am about to fund it on Monday.

Thank you. I look forward to hearing from you.

Jim

[judgetrade]

Great work Jim, Kudos!
one question: what kind of CAP size does the above model pick (small, mid big) and in what percentage. Thank you!

Andreas

[taofen]

Thanks Jim for sharing this. I believe many of P123 users are interested in your sharing/findings while we are waiting for P123’s ML feature coming out. During past year, i also spent my spare time on this machine learning with P123 data, got some progress but no success yet. If the ML is also useful for mid or large cap, that will be super. Thanks

Taofen

[Jrinne]

Andreas,

Thank you.

It is the easy to trade universe. In my ranking system there are no explicit market-cap factors (e.g., I do not use market-cap smaller is better). And market-cap is not in any formula so, I think, the ranking system should no be selecting stock based on their market cap.

I am never sure if this covers the entire history or of this is just the most recent holdings but this may help some:

Jim

[Jrinne]

@taofen

If the ML is also useful for mid or large cap, that will be super. Thanks

This works for SP500 universe with not a great drawdown (i.e., more drawdown than anyone would want). I believe in diversifying. I would not mind making it a designer model. Would you want this as a designer model? If so, give me your honest opinion of the right price and I will just do it.

Jim

[Jrinne]

BTW, I am following myself in my own thread so I believe I may be safe. Would this be considered a taunt?

Feature request. Emojis contain emotional content which will almost certainly be offensive to someone. Please remove the emojis. Not a taunt but a real request.

I was kind of a class clown. They never sent me to dentition because Robert K. (my partner in what ini retrospect was stupid juvenile humor that probably had more to do with hormones and making the girls laugh than good humor) was the smartest person in the school (a math professor now). One of my fondest memories from high school was Rob and I sitting in a literature class and discussing a short story. The whole point of the story was that, according to the teacher, it made a “point thru irony.” Rob of course shared with the class that this was really cool because he thought he was always just being sarcastic.

But also humor and much true intellectual progress is based on reductio ad absurdum or an indirect proof. And maybe I make too much of this but I truly believe his use of reductio ad absurdum and nimble us of it at that is a large part of why he loves math, was so good at it and the smartest person in the school. Despite its usefulness in understanding things in a serious way, there is a large percentage of the population who become offended with any indirect proof (serious or not). It is the source of all humor and like music it just engenders an emotional response this is not fully understood, in my opinion.

BTW, I have read many true works about fiction on Hamlet that discuss his use of satire and what that says about his character. Rather than read those books anymore i just ask ChatGPT. Here is just a small part about Hamlett: "His soliloquies, while deeply introspective and often melancholic, contain moments of self-directed irony."

But my mind uses it much of the time (not as well as Hamlet I might add). When it works as a proof it is the most compact and air-tight proof there is. And as Rob would say: “Maybe I am really just a sarcastic dick in the end.”

Rob’s humor could have an edge at times but only when someone was messing with him. He was/is kind.

Anyway, to the best of my knowledge my email provider has not threatened to remove me from their system for me making a point thru irony. You can reach me there about my work. I will continue to post very serious questions about programming and technical issues (with no emojis: seriously please remove them). I pay for that I believe,

Jim

[Jrinne]

@Whycliffes

All, I went ahead and paid the $500 dollars for API credits. Plenty of credits for what I needed. Here is a k-fold validation with a test sample and embargo period of 3 years (pretty aggressive embargo period) comparing recursive feature elimination and no recursive feature elimination. As you can see the results are similar. I have done other analyses. [K-fold Valdiation] But I did a k-fold validation on all of the data after this with no recursive feature elimination (also no test and e…

Hi Jim,

Are you willing to share your ranking system from this test?

You wrote in your post that P123 will soon come with machine learning. I have heard this for some months; do we know anything about when the feature will be released? And do you know if this will be very difficult to use?

This is from my email. I think Whycliffes was emailing as I requestioned and would not mind me sharing here.

@Whycliffes

Are you willing to share your ranking system from this test?

No.

@Whycliffes

You wrote in your post that P123 will soon come with machine learning. I have heard this for some months; do we know anything about when the feature will be released? And do you know if this will be very difficult to use?

Based on everything P123 has done so far I think it will be excellent and easy to use.

So to digress just a bit, I do not think people always realize what a cool thing the ranking system is with its option of composite nodes etc. AND THE FACT THAT—WITH THE USE OF THE OPTIMIZER—IT IS A MACHINE LEARNING ALGORITHM. That and other things (too numerous to mention) is what I mean by “everything P123 has done so far.”

I have almost nothing in the way of additional information. I think Marco posted that we will be “able to incorporate it into the ranking system.” That is not an exact quote but based on memory.

So you can interpret the “quote” without me but I think it suggests excellent ease of use.

TL;DR: I am sure Marco et. al. have done a great job here. He always does. Why would this be different?

Jim

[taofen]

Thanks Jim for posting this SP500 result. I may not be interested in subscribing the model as of now, but i think you can try to make it as a designer model, when its oos performance is good after half a year or more time, there will be other people interested in it.
I just want to say your result is inspiring us including P123 to continue to explore ML. Thank you for your contribution to this community.

Taofen

[pitmaster]

Hi Jrinne,

Thanks for sharing your system.

I try to understand high level concept of this framework, and in general how to leverage ML and create a P123-compatible ranking. Input for P123 ranking (which is your ML output) can be only in a linear form.
Therefore the output (ranking, R) of your system also needs to have a linear form. For example the first factor (f1) is f1 = (w11 * ratio1 /ratio2 + w12 * ratio3), f2 = (w21 * ratio1 ...) … then ranking R = (w_f1 * f1 + w_f2 * f2 + ...).

So, you may perform trillions of combinations in python. You may use brute force optimisation and/or ridge/lasso regression to find factors and its weights. Then you can cross validate your selected model (factors) on test samples. And finally your linear ranking R is feed into P123 ranking system and you test your results here on the platform.

And the questions are as follows:

• Is this more or less the framework you follow ?
• Is there any moment in your framework when you use non-linear/advanced models/techniques (e.g., for feature selection) ?
• Are the factors’ parameters have expected sign in your model ? (e.g., higher Sales/EV → higher predicted return), or do you have no constraints on this matter ?
• What is unique (in general) in your framework that allows you to think that your model is better than a standard ML-based stock selection models (no offence, just curious)

Thanks, Piotr

[Jrinne]

Piotr,

I have been banned from P123 for this post according to an email. You probably will not receive this. If you do I am sure the glitches preventing P123 from stopping my posts will be fixed shortly.

Sorry, I will not be able to answer your great question until there is some sort of final resolution on the value of machine learning posts to P123 (with or without some irony and/or humor)! Honestly, I do not think the occasion joke is the problem. Some people have strong bliefs about what works. I will say I think Marco had invested a lot in machine learning. I am not questioning his belief in machine learning for a microsecond.

And I cannot think of better way to say thank you to Marco for that belief than my last thread (probably) here.

Jim

[Jrinne]

BTW, I have learned to screenshot things so that I am not later told:: “I have no idea what you are talking about.”

TL;DR: Posting even the most positive post for P123 ever written, IMHO, is dangerous.

[judgetrade]

thank you!