What a shitshow – https://news.ycombinator.com/item?id=46154491. – HN midwits are outdoing themselves in trying to signal their virtues and unique snowflakeness.

Turns out I know enough math, programming, LLM theory (including the underlying math) and trading (in the Elder’s tradition) to understand that this is an utter crap – “Everyone is a genius in a bull market”.

This is bullshit at so many levels, and yet, someone paid for this and the “researchers” got the highest possible “living standards”, of course.

First of all, a general-purpose LLM could not trade in principle. It can be made to appear as it does, but an appearance is usually very different from what is actually going on – a parrot appears to speak a human language (while it just mimics the sounds), dogs appears to understand a human language, while they just respond to a familiar sounds, “read” the tones of a human voice and use non-verbal cues.

Here is how. The first principle is the classic one “The past performance is not indicative for a future one”. The model which has been trained on the “past” cannot in principle operate efficiently within an already changed environment. Yes, one could “drive” the road using only its memories, but only till the next obstacle. No, all the self-diving cars have tons of real-time sensors and process the data in a real time, combining the sensory input with the model’s representation (sort of mimicking our current understanding of how biological systems work), which is NOT what the general purpose LLM do.

Second, betting on tech stocks (even just the blue chips) in the unprecedented, of an astronomical magnitude tech bubble, with NVDA going parabolic, is not an indicator of any kind of intelligence. It is not even a luck, it is just the ability to buy.

Every interpretation they gave to the observed performance of these LLMs is simply, in principle, wrong. The difference in scores is due to chance factor, not any kind of predetermined abilities or even differences in training of said LLMs.

The experiment can be dead simple – simulate the same setup with the recorded (and now historical) data and measure the result – there will be no correlations whatsoever, and one does not have to be a rocket scientist to understand precisely why.

Last but not least, even training (and backtesting) a sophisticated classifier will fail (but, perhaps, will outperform a random-walk strategy) because, again, any market data is “outdated” the moment it released (observed) and the “environment” is quickly “evolving” right behind your back.

There are reasons why HFT is considered the only “marginally or mostly deterministic” approach, since the “patterns” they are trying to capture and react to are “stable enough” for just a few moments (until the others react to them).

The only non-bullshit trading strategies (aside from inside trading, of course) is to trade the tops and the bottoms, but timing them is very costly and very stressful, and, of course, “the market could remain irrational longer that you can remain solvent”. Even Michael Burry has thrown the towel in trying to time the market (TSLA and NVDA obvious bubbles). “Time in the market beats timing the market” is the most fundamental principle, along with “buy only when there is blood on the streets”. Everything else is just bullshit and memes.

By the way, semi-automated trading, using a most straightforward classifier on candle-stick and some indicators (BOLL, RSI) patterns will outperform most of the strategies if (and only if) it is used to time a top (or a bottom) of a market cycle, which is, again, very costly due to constantly being stopped out or liquidated by scam-wicks.

Trading in the middle of a range is just gambling by definition, and no, your LLMs aren’t any good at it.