Hedge-Fund Review Highlights Weak Spots in LLM Stock Forecasting

If a trading model looks smart in a backtest, it can still fall apart when real money hits the market. This review looks at large language models (LLMs, AI systems trained on huge amounts of text) in stock price forecasting through a hedge-fund lens. It brings together recent uses of LLMs for reading financial news and social media sentiment, analyzing financial reports and earnings-call transcripts, turning price series into tokens or symbols, and building multi-agent trading systems. The paper also flags the traps that can make results look better than they are: fragile sentiment analysis, poor dataset and horizon design, weak performance metrics, data leakage, illiquidity premia, and the basic limits of stock price predictability. The point is not just to show what LLMs can do in finance, but to stress-test whether they can survive the messiness of real trading pipelines.

A model can sound brilliant and still lose money. That is the warning at the heart of this review. It looks at large language models, or LLMs, which are AI systems trained on huge amounts of text. In stock forecasting, they can read news, social posts, reports, and earnings calls. That sounds powerful. It also sounds dangerous. A trading desk cares less about polished demos than about what survives slippage, thin trading, and bad data. This review, from a hedge-fund lens, asks a simple question: which LLM ideas still work when the market stops being neat?

Where LLMs fit into stock forecasts

The review groups the field into four main uses. First, LLMs pull sentiment, meaning the mood or tone, from financial news and social media. Second, they read financial reports and earnings-call transcripts for facts and clues. Third, they turn price series into tokens or symbols, which are small text-like units a model can handle. Fourth, they sit inside multi-agent trading systems, where several AI parts work together. That spread matters because stock forecasting is not one task. It is a stack of tasks. Each layer can help, but each layer can also fail in a new way.

The traps that make a backtest look better than it is

The sharpest part of the review is its caution list. Sentiment can be fragile, because a positive-sounding post is not always bullish in a market sense. Dataset design matters, because the wrong time window can hide the real test. The choice of forecast horizon matters too, since predicting the next minute is not the same as predicting the next month. Evaluation metrics matter, because one score can flatter a model while missing the real cost of bad trades. Data leakage matters as well. That is when future information slips into training by mistake. Illiquidity premia, or extra return tied to hard-to-trade stocks, can also fool a model into seeming smarter than it is.

Why hedge-fund readers should care

A hedge fund does not live in a clean notebook. It lives in a market with costs, delays, and noise. That is why this review keeps returning to robustness, meaning whether a result still holds when conditions get ugly. An LLM that spots tone in a headline may still fail when the headline is vague, sarcastic, or stale. A price model that shines on one dataset may collapse on another horizon. A multi-agent system may look elegant on paper and still stumble in live use. The review does not reject LLMs. It asks for harder tests before anyone trusts them with real capital.

What this leaves for the next test

The next test is not another neat backtest. It is a real trading pipeline with market friction built in from the start. The review points straight at stress tests for data leakage, illiquidity, and weak sentiment signals. It also points at the basic fact that some stock moves may never be very easy to predict. That is the surprise here. LLMs are not just text machines. In finance, they are also reality checks. They show where a model can read the market. They also show where the market refuses to be read.