Good questions. It really comes down to two things: what do we mean by AI/ML and what data are we using to answer what question.
If we mean deep learning and very large models, those make sense only when we have stupid amounts of very high dimensional data. Think all of the internet (what ChatGPT is trained on). Models like that are crazy expensive both to train and to run. But letās say we make them more economical. Does it make sense to use such a model for trading? I donāt know, I suppose the latest news articles will affect the market short term, but does it make sense to exaggerate daily bull and bear responses with ML? All of this leads to a feedback loop where the responses of the AI make it even more likely to respond like that in the future (buying something makes the price go up, etc). Somehow it doesnāt make sense to me, especially long term. But I suppose very short term, you are right: if only one player has a good large model for predicting daily trends from news articles, they may be able to do more successful day trading. Until others catch up. If this happens, I predict the market as a whole will become super crazy with lots of volatility for some time.
If you mean more established statistical and ML methods, those are already used all across finance. Iām not super familiar with the nuances here, but all sorts of regressions, time series modelling, probabilistic programming is definitely used, both for trading and for fundamentals / longer term strategy analysis.
I definitely donāt think the market is too complex for an AI to understand. But how we use data to answer questions is complex. Itās so easy to build a nonsense model that predicts nonsense but looks like itās working. It reminds me of an example with the police in some city in the states (donāt quite remember which one). They build a model to āpredict areas in the city where crime is likely to occurā. They trained it on data about where historically crime has been observed to occur: in other words, areas where the police was patrolling and observed crime. Of course, the police is more likely to patrol certain areas (e.g. black neighbourhoods) than others, so of course it will observe more crime there than in other areas. Having been trained and retrained on this data, the AI simply predicted more and more and more patrolling in those areas, which led to more and more policing of these areas. I hope itās clear that what we wanted to predict is crime, but what we actually ended up predicting is where the police is patrolling. And on top of that we created a positive feedback loop. Something similar will happen if we implement trading AIs.