Good one, happy to add my perspective here:<p>DISCLAIMER: I've spent the last 8 month heavily on building a quant-based asset management app (though, still not live, currently in final steps to sync processes with broker)<p>a) I tried to leverage some of this AI-voodoo stuff, though not on the level as in the paper; my findings are clear (at least for me): AI-driven trading does not give you a bigger/better edge than any of the other well-known approaches<p>b) In fact, AI-based approaches are at best on par with traditional approaches, in lot of scenarios not even this; I havent seen any setup from anyone which actually outperformed one of the classic approaches.
BUT: The AI-guys have much higher cost, be it Infra, processing time / waiting time in front of screen etc.
So you have you to pick carefully, which one you choose.<p>c) I'm doing today only "standard approaches" with volume/statistics/vola/price action, as this approach is super-cost-efficient (i need only one cheap datastream) and a lightweight machine for 10 / 20 USD a month<p>d) It is clearly possible to outperform the market, though these approaches are not scalable unlimited - Ex: depending on the used instruments, there may not be enough liquidity to buy continuously for 100k, but maybe for 10k only.
Apply leverage of 5-10 on an asset that moved 5% in last 10 days on a 10k position - is this outperforming? A clear >yes< in my perception?<p>e) People who have built & found a stable approach do not share it or talk about it, there is no real community; you will get details of working approaches only from people whom you are really "friend with"; there is a lot of unshared but working business tactics in the field.
So they used a LLM with knowledge cut in mid 2023 to evaluate 2023? Seems like a classic leakage problem.<p>From paper: "testing set: January 1, 2023, to December 31, 2023"<p>From the Llama 2 doc: "(...) some tuning data is more recent, up to July 2023."
> Alpha Factors incorporates 108 technical indicators and factors with their expressions,
which are believed to possess predictive power regarding stock price movements.<p>Examples of the indicators are in Figure 15. The ablation studies in Table 4 suggest that market and news information made a much bigger impact than the magic indicators. Makes sense if the indicators are simple enough that the LLM can reproduce them without losing processing power.<p>I somewhat like that they used DJI and not SPX, but 2023 was a sideways bull year with DJI +12% and SPX +23%. One year is way too short of a study.<p>> Hardware: NVIDIA A5000 GPU x 4, AMD Ryzen Threadripper PRO 3975WX CPU,
256 GB RAM<p>Seems approachable.<p>> The proposed TradExpert framework utilizes a Mixture of Experts (MoE) approach, where four LLMs are specialized in processing distinct sources of financial data. All these LLMs are based on the LLaMA-2-7B Touvron et al. (2023b) model and fine-tuned using the LoRA mechanism Hu et al. (2022)<p>Relatively small LLM.<p>Overall, this does seem like an interesting study, even for just comparing data sources.
If I understand this correctly we have come full circle on what MoE means.<p>MoE started out as some form of multi model approach.<p>Afaik in current architectures it’s basically a load balancing method that while it increases latency makes the model better suitable for distributed operations.<p>To me this reads as if the author uses the term closer to Urs original meaning than its current.
How do people on HN think about the market?<p>Do you think the market is so efficient that anyone who outperforms it is merely lucky?<p>Or do you think the market is inefficient enough for a person smart enough to be able to outperform it by thinking?<p>In other words: Do you think a single person can rationally decide to invest their time into thinking about the stock market? Or would that always be a fallacy, and whatever the outcome is - we can't decide if it was just good or bad luck?
Does anyone understand how the Market Expert works? It takes in numerical OHLC data and converts it to embeddings for use by the LLM… but embedding are also numbers so I don’t see how that’s any easier for the LLM to process since it’s a language model.<p>> The Market Analyst LLM focuses on analyzing historical OHLCV (Open, High, Low, Close, Vol-
ume) data to predict stock movements. However, time series data is inherently continuous and lacks
the discrete token structure that LLMs are designed to process. This misalignment poses a signifi-
cant challenge in effectively utilizing LLMs on time series. To this end, we utilize a reprogramming
mechanism Jin et al. (2024) to reprogram the input financial time series into text prototype repre-
sentations.
The market is like an ecosystem. There are huge mammals (investment banks, hedge funds) that look at certain type of preys, and there are smaller rodents that only eat very tiny worms.<p>In high volatility regimes, ie. stocks with low market cap, the market is far from efficient. Hedge funds are not even looking at stocks with 100M market cap.<p>There are traders that act in these regimes that beat the market, exactly because they play small.<p>Anyhow most people would be better of by assuming the market is completely efficient.
fwiw, I tried something similar about 5–10 years ago. I wasn’t using LLMs like the abstract here suggests, and honestly, I’m not sure how you'd act on a signal fast enough with them. When I gave it a shot, there was some slight predictive value, but in the end it felt like noise and gambling, so I moved on.
How did it perform against a boglehead portfolio? Were fees and commissions included? Seems weird to evaluate performance over a single year for trades. Much more interested in long-term growth over one or more market cycles.
Not a very good study. I did not look at any of the researchers background but it’s like they did not consult their respective finance school departments.
I'm curious at what point stock and derivatives trading becomes something entirely for AI, and thus restricted to companies rich enough to buy tons of GPUs. Are we near that point yet?