Yes, the $50 credit is a one time deal. It gives users an opportunity to try AI Factor at no cost to them.
1 Like
We'll have a 50% discount for the first 6 months of the prediction add-on by next week.
Whoever purchased the add-on will get a refund, or the next bill moved a month later.
Thanks for the feedback.
3 Likes
Thanks marco! This is appreciated.
Since machine learning uses relatively rigorous out-of-sample validation, many people find that their validation results are not as good as their backtests and assume that machine learning is useless. In fact, this just tells them that out-of-sample shouldn't be expected to be that good in the first place.
In fact, the linear model does a much better job of fitting than people do on their own by repeatedly modifying the ranking system and backtesting. The downside is the lack of shrinkage, but people's own fits are even worse.
2 Likes
I partially agree with you, but I also believe there is some skepticism because this is a completely new method and approach compared to what most members have used so far. This means that, along with others, I am likely waiting until there is a more comprehensive description and guidance on how to use ML on the platform the best way, and I can see that more people are starting to publish their results by using the models.
I had some ML models with very good returns in the ML 'results,' but for some models when I tested them in the simulator, they became extremely volatile, and the excess returns could be attributed to just one or two years. I also see that there are several forum posts pointing this out.
If I had a wish before I start paying for such a service, it would be for a more extensive manual that includes several recommendations on settings and so on...
I know I can stand do learn more. And I hope that never stops. I guess I wish the forum was more curious in an academic sense myself. I can wish for a lot of things but I am ultimately responsible for what I do with my investing.
I think P123 has done an incredible job of providing an almost automatic solution.
The cross-validation is excellent and almost foolproof. In fact, it might be entirely foolproof: you're unable to run a simulation with in-sample data. That's pretty hardcore (meant in a good way). Maybe you have to understand not to use training data in in the validation data when you set the gap.
Some of the P123 ML models (Like Extra Trees Regressor) work unmodified. Have fun and change some hyperpaprameteers if you want, but you don't need to. Did Dan do some of that along with Marco and Richard? Good job whoever did it.
We can all learn or do more. And I hope that never stops.
Stepping back and getting the big picture, P123 is competing against itself a little (API vs ML) and I am not sure how P123 should work that out to everyone's benefit. The goal should be for all of us to win (and profit).
My recommendation, for what it is worth, is to use the excellent ML and data offerings to grow the membership and put less emphasis on getting more money from present members. Maybe bring the costs down with more efficient algorithms (where possible) and economies of scale.
The AUM (asset under management) of many members just does not support some of what present members would like to purchase no matter how good a new offering may be (previously pointed out in this thread by another member and true for me too). I do diversify with ETFs and use Portfolio Visualizer for that which, of course, affects my ratio of cost to AUM ratio calculations—even if I have nothing to complain or whine about as far as total AUM.
But what to do with all of that is a business decision and I am just a mathematician hobbyist. Glad to have some data to play with and use for making my decision-making more rational. I thank P123 for helping me with that.
Despite my lack of formal training in business, I know enough to see that the combination (formal ML and downloads combined with ML/algorithms) is a pretty incredible offering. I will be using one or the other or both.
Wait, didn't you guys only invest in p123's advice?
I just think that if you think you have an advantage, of course you should go all out.
I was also wondering why my results were so different from yours.
Is this directed at me? I think you may be asking a great question about rational AUM /investment cost ratios as well as different ways of balancing risk that I could learn from but not sure I understand.
For sure Portfolio Visualizer and ETFs is not the only way to manage risk and probably not the best. I concede that.
FWIW, I use a risk parity strategy that includes my ports in the risk-parity calculation and I overweight the P123 ports based on expected outperformance. My P123 strategy adding to the diversification with its overweighting in small-caps and perhaps value.
I am near retirement and my wife would probably divorce me if I told her about a 50% drawdown adding to my overall decision-making as irrational as she is. 
2 Likes
I mean I think people here don't invest in other things than P123 model's advice.
Probably true in general. I agree.
So for a young newbie willing to take a 50% drawdown and a long-term view (but possibly having no idea what she is doing yet as it the case with some Designer models) what is your recommended AUM/cost ratio? Serious question. I will leave it open and not follow with any "Yes, but…. " lame remarks on my part unless you ask for my input on that. Open question (as I know you will have a well-reasoned answer) that I will have no reason to debate anyway. And emphasize again, my wife's irrational view on temporary drawdowns is not to be ignored for me personally. Not necessarily a joke above although probably (maybe) an exaggeration. In any case, it is her money/retirement too (although technically my SEP-IRA, I guess, in a legal sense).
But my question is about newbies which is not personal for me.
Thanks, we are trying to stop people shooting themselves in the foot. Lots of work to do still. P123 is still a dangerous place, specially with the older tools.
The web based, all in one ML product, with no coding necessary will continue to evolve. We'll see what happens.
We've been at this a long time. Time to face the facts. Our original "mission" to provide professional tools to the little guy is untenable. Several startups that tried this have failed. Retail users that can use P123 is a super small niche with lots of churn. We have to target non retail. To target both is tricky: if we price retail much cheaper we'll end up with a bunch of pros registered as retail.
4 Likes
The web-based model is pretty standard-fare for now.
People like Pitmaster, ZGWZ, Bob, Jlittleton, SteveA, Wycliffes, P123 staff now and too many other members for me to think about at the moment, have been generous with their personal ML methods and insights.
@yuvaltaylor has always been generous as far as feature engineering and his own personal optimization method. I believe it is his preference not to call his optimization method machine learning but he has contributed a lot.
Yuval has himself to protect as well as an obligation to his hedge fund, I believe. Maybe other professional members too? Or people who, in essence, have their own personal prop-shop
Obviously, when methods are incorporated into the web-based AI the "cat will be out of the bag." But that can be a long time and perhaps never for some good ideas.
Is it best that we share everything with every retail investor and professional person on the planet thru the forum going forward? Will that stifle cooperation and development of personal models as well as P123's AI?
Since we are not appealing to retail investors so much anyway? Is the entire forum public? Should it be if it be if it is in fact all public? Seem i have accessed the forum at the hopital without logging in giving me a potentially wrong impression about this.
P123 does have the potential now to be a beginning-to-end professional system. Or there is the potential if not quite true for some now. Not every hedge fund talks about their strategies on the web.
Standard ML 101 being an exception. People fool themselves if they think that is new.
I don't mean to imply this is a priority of mine or that this needs immediate attention. But that same irrational wife I mentioned above keeps telling me to "shut-up." For now, I learn more than I lose in any duplication of my strategies, I think. Maybe, considering I use exactly zero P123 AI/ML methods for my presently funded,….uh, just an algorithm using some Sklearn programs, I guess. Plus, ML 101 is no secret.
"Yes, anyway, thank you for the help! I received a lot of input from you that was interesting for my learning.
Regarding the ML models, I had many good results. Below are some examples where there are around 30-60 stocks in each bucket, but with slightly different universes:"
Micro Universe:
"Again, I got plenty of good results, but for some reason, I couldn't achieve any near-similar returns when I tested it OUS in the simulator. Or, I had good results in the simulator, but they were very volatile."
1 Like
I don't know, but I haven't seen a spike in trading volume in the stocks I'm invested in. I guess my secret isn't out.
Yea, Using LightGBM? If not you want to share? List of features if you have no concerns?
Not interested in sharing my ML model or features myself. Just a choice. Many share exactly zero about what they do or even much of their general knowledge for whatever reason(s).
I appreciate what you do share no matter what else you may wish to share. You share more than most and have helpful advice.
Now anyone can read what you say (P123 member or not), right? My question: A good thing?
The list of features for the system I currently use comes entirely from dan's sheet. I always said that.
1 Like
Have you tried those linear models? I think they are more stable.
I have. As I said I use nothing in P123 offering now but I am looking into elastic-net regression (a not-so-secret algorithm that I have no problem sharing).
Also, I was not aware that P123's linear regression was actually elastic-net regression until you told me. Thank you for that. I am using downloads for this and I cannot access P123's elastic-net model to add to the information you gave me on that.
Thank you! Good advice, I believe.
I get this, and personally, I'm willing to pay more for add-on features if I have a reasonable confidence that those features add positive expected value (net of their costs) to my trading. I've already spent quite a bit of time to develop some model-building infrastructure using the API, so whether the ML functionality is +EV for me remains to be seen -- though I'll bet it will be.
To help drive adoption of the ML add-on, I'd suggest one modification to the pricing model. You currently charge $100-200/month for using the AIFactor prediction function in screens and backtests (I'm assuming AIFactorValidation falls in the same boat, though the pop-up doesn't make this clear). Given that some users have already observed some of these AIFactor lift charts not directly translating to backtest performance, users may be reticent to commit to a big jump in recurring costs just to look at in-sample backtested performance, even if you can cancel the prediction function add-on at the end of the month.
Instead, what if you allowed AIFactor/AIFactorValidation() free for users except for say the most recent year, or potentially the most recent period of their target horizon (e.g. the most recent month if their target variable is 1MRel). This would greatly lower the barrier to entry for experimentation and could end up with more paying customers in the long run. Since we can't trade off the predictions in the free window, users would have to pay to use them for live trading, but we would have a much better sense if they will be +EV first.