IMO, the problem with PRT is its website through PRC and the withholding of extensive discussions on “work outs” of scenarios that an LLM would learn from (or better : would rely on). This goes deeper than even back intended, but it is Market Place related why this now does not work at all (the discussions ceased say 4 years ago). A huge mistake for PRT itself, or even worse : the worst mistake that PRT does not own a web site itself and instead depends on the instance ProRealCode (which is 3rd party – may you not know), with commercial interest.
Thus, this damage has been done and besides, it will be too tough to ever revert – PRT themselves will be interested the least. Oh, they will acknowledge in maybe a couple of years, unless they keep their edge as being a 100% server side platform, which is unique (AFAIK) (read : they may keep on selling anyway).
I must honestly say that I have few to no experience at this moment in SLM, so I would not know what that could bring, but as far as I can see this would require training which “we” are not going to accomplish. This put differently : this training could theoretically be localized by You or Me or other individuals (or Groups ?) which is highly inefficient, related to LLMs which would dig out merits “by nature”. … Too much BS from me ? just correct me and ready to learn !
An advice could be to take your own learning and applications to a higher level by not using PRT as coding language at all. Yes, this means that you must be capable of dealing with ProBuilder and the Backtest/ProOrder subset, but hey … doable. You could order for natural language (or any language you know *and* is known to the LLM) which you can take for input to your own PRT code. This would be the least problem – says me.
Let me try to give you an exciting example :
Only last week I tested ChatGPT (4.5) for its capabilities of determining changed markets since Trump, Tariffs in particular, and more. But, I added the dimension of my own – mind you – AutoTrading results, just showing the equity curve per date or any detail 4.5 would ask me for. From that came the 100% accurate analysis of what my code does for Entries – which I actually asked for, heading the way for improvement.
100%.
For me this is long gone not astonishing, hence expected and thus usable. It is not related to details of the language at all, which btw would be highly different from PRT-IG vs PRT-IB. Thus, this devil-in-the-detail would not be worked out by a LLM anyway because you and me would be too unique to begin with, and I (nor you) will not discuss in public what I all do to have this technical edge. It would all over fail (which is your subject, NT) because it would, for example, impose a Trailing Stop at 0.5 points for IG which can not exist, or a Trailing Stop at 0.5 points for IB adjusted each 1 second (TF) which would be allowed but only can exist per your own code instead of allowed commands for IG, but which will fail because of IB choking at the change each 1 second which IG would bear. Bzzz.
GPT LLM would rely on posts like this times 1000 in order to help you out with that. Without this, the SLM would rely on your own knowledge about this all, which cannot exist (too much knowledge required). Back to PRT itself … there it could exist, but also does not (trust me). Then back to this PRC website/forum … nobody (owners) care.
All in other words, your subject is great and probably only understood by you yourself (and a handful of others who may persist, being in the self-learning phase (I have that behind me)). It is a dimension new to the world of consumers, and once detected actually too late to cover for. For fun : ask GPT (just 4o) how to solve this – hence what to undertake to solve this. You will soon learn that it is too late for this decade (this is my idea about it, but try it – best is to leave out the suggestiveness).
One last more – to the point – response : you may not know it, but PDFs exist for ProOrder/ProBacktest. This is actually what you asked for, but it would only be my very last response. As you can see. 🙂
Please talk back because I am very eager to learn myself.
Peter
Well Peter,
That is indeed a very much appreciated heads up of the state of PRT affairs, and its fundamentals – I thank you for making the situation, and the foundation of PRT Platform as clear as possible.
The fundamental challenge using AIM like ChatGPT, etc, is the historical context. Most if not all, even with performance updates are preloaded with data some 2 to 3 years previous. Even if the AIM can access the internet, there is still the case of code segregation of that which is for the ProBacktest module.
As for training an AIM SLM/LLM, is a lot simpler than of late. Take a peek at Langflow or even Flowise (Pref).
These GUI interface design platforms, make the building of a learning AI more likely (That is the AIM is fully aware of said system rules, and its interaction with code glossaries and the like).
The main issue, in this case, is the differentiation of the code related to ProBacktest (PBT) and the ProBuilder Code (PBC).
If, or once the code glossary and supplemented examples are implemented and relating only to the PBT, and if the AIM is subjected to reinforcement learning (RL), it would indeed be taught to seek its own self approval of correct or error based code, or indeed a manual method of RL is applied:
- Request the AIM SLM to define said code (Using defined instruction templates).
- User applies and tests said code on the PRT Platform.
- User reports back to AIM SLM, using a defined feedback loop; As an example:
- +1 = Well done, code is working as defined.
- -1 = Code fail with explicit reasoning (AIM would store and relate to this memory during each coding session).
- 0 = Working, but the output is not strictly correct.
The issue isn’t in working with or building the AIM, the fundamental of a AIM SLM, is its processing of data: known as garbage in = garbage out or good intel in = good intel out.
The AIM working model, and its ability to produce the correct code output, (That is for the PBT) is all bound by ‘feeding’ the AIM with the correct code in the first place.
Hence, my post to ascertain such code.
The question is, using an AIM SLM, which is coded to learn (Machine Learning > ML (Python)); just how far can the current code be pushed, to produce backtests, live order processes, that are ‘out of the box’ thinking?
More so, to keep, or to ensure this ‘Trading Product’ is at the edge of its game. The application of AI, especially AIM ML is essential to its progress, as many coders are discovering.
ML AIM have made promising out bound steps, where we are shown potential avenues of change, of engagement to the betterment of outputs and the utilisation of said foundation code requirements.
I do so hope, that someone within the realms of the PRT Platform, can produce and share such code that is only for the ProBacktest module, as this would indeed open up so many possibilities for the platform, the code, and all of its users, established or new.
NT.
