AI hype, from Nils Poertner

September 18, 2023 |

remember the hype about Chat GPT some weeks /months ago? def for trading /investing - I doubt using that or any other program will help to master time ahead - prob a recipe for disaster at the end.

Peter Ringel writes:

I am still hyped! Hyped for boost in efficiency of the economy via AI. Not hyped for AI-trading systems! So far the training data set seem too small for AI - trading, thankfully. Together with what the Senator and others posted here: humans still beat skynet. Yet, I like to remind myself every day: the bastards are coming.

Hernan Avella responds:

So far the training data set seem too small for AI - trading , thankfully.

How do you figure this? Each trading day probably produces more than 100's million rows between trades and quote updates for all levels and exchanges, if you include futures, equities. I don't think lack of data is the issue here.

Peter Ringel replies:

I know even less about AI-coding, than about trading-coding. So everything is based on perceived experts. Thankfully, so far they are pessimistic.

Hernan Avella continues:

So everything is based on perceived experts.

The set of experts in ML-DL is very small, and the set of experts in trading is also small. I imagine the intersection is even smaller and more importantly, secretive. My suspicion is that the training set is more than enough, but the problem of ergodicity and stationarity (lack of) of the ever evolving competition are the culprit.

Peter Ringel responds:

I hope, you are wrong with this. But at some point you will be not. I speculate, that the "small" existing universe of trading history data + some sort of data - > model on human psychology - will be enough - will make us traders obsolete.

Peter Saint-Andre writes:

In my limited, non-trading experience with LLMs, I've found that their output reflects conventional wisdom. That might leave plenty of room for creative strategies outside the mainstream.

Peter Ringel agrees:

yes, they are regression x1000 on speed. so far feedback loops/ "reflexivity" kill it. As far as I understand.

Hernan Avella warns:

I would abstain from making any statements about the state of the art ML applied to trading, specially from a place of ignorance. Whoever works in this field (which there are only a handful in this list), and interacts with just the basic chat GPT 4.0, realizes immediately the productivity boost and immense potential to improve one's process. Only a moron would expect a good output from just feeding prices to the engine or asking simple questions.

Peter Ringel agrees again:

nooo! especially if you are ignorant in a field , better check if that poses a risk to your systems. I believe AI is a risk to traders. Here is a fact already reality: ChatGPT empowers people to do substantial back-tests.

Big Al adds:

And doing backtests poorly, or being improperly overconfident in backtests, is a threat to one's trading.

Humbert K. wonders:

With reference to the skynet, it is hard to guess if and when fully autonomous weapons will happen. My 2 cents is: Fully autonomous weapons will happen. There are debates as to whether we should let machines make kill decisions. I can say though our adversaries' weapons developments will not be bound in any way by any moral or ethical standards. If the bots can communicate with each other and collaborate to perform. When will they no longer need human inputs or interventions?

Eric Lindell writes:

There's a limit to what computers can do with the massive amounts of data available in countless categories. To find the perfect mix of factors to plug into a formula — if there is such a thing — would require a number of operations that increases exponentially with the data-set size.

Humans are good at intuitively navigating such complex search spaces. Computers using brute force just aren't powerful enough yet — and may (in principle) never be. That said, if a human comes up with a plausible conjecture relating stock picks with subsequent price performance, computers can certainly back-check the theory.

I'm working on one now regarding immediate post-IPO performance of stocks selected by certain criteria — criteria that aren't widely (or even narrowly) recognized for their relevance — pertaining to historical research of a revisionist nature.

Recent Posts

List of Authors

Sep

18