Efficient Market Hypothesis (EMH) and the Sentiment Frontier

The Efficient Market Hypothesis (EMH) posits that asset prices fully reflect all available information. If markets are perfectly efficient, it is impossible to consistently achieve risk-adjusted excess returns ("Alpha"). In the modern era, the "frontier" of efficiency is defined by how quickly the market processes unstructured data—news, social media, and transcripts—using Large Language Models (LLMs).

1. The Three Forms of EMH

Developed by Eugene Fama in the 1960s, the EMH is typically categorized into three levels of information integration:

- **Weak-Form Efficiency:** Prices reflect all historical trading data (past prices and volumes). Technical analysis (charting) cannot generate alpha.

- **Semi-Strong Form Efficiency:** Prices reflect all *publicly available* information (earnings, news, macro data). Fundamental analysis cannot generate alpha because the market reacts instantly to announcements.

- **Strong-Form Efficiency:** Prices reflect all information, including *private/insider* information. Even insiders cannot achieve excess returns. (Generally rejected in practice).

2. Sentiment as "Noise": The LLM Impact

The **Noise Trader Theory** suggests that some investors trade on "noise" (sentiment, rumors, or irrational exuberance) rather than fundamentals. Today, LLMs act as both the generators and filters of this noise.

The Sentiment Frontier

LLMs have pushed the Semi-Strong frontier to the millisecond level.

- **The Loop:** An earnings transcript is released $\rightarrow$ LLMs parse the sentiment and extract key KPIs in <100ms $\rightarrow$ Algorithmic traders execute based on the LLM output.

- **The "Noise" Injection:** LLMs can also introduce noise. If an LLM misinterprets a "sarcastic" tweet or a nuanced CEO comment as "highly bullish," it can trigger a flash-spike that is decoupled from fundamental value.

Concrete Example: LLM-Driven "Flash Sentiment"

Consider a pharmaceutical stock.

1. **Event:** A clinical trial result is posted on a medical portal.

2. **LLM Noise:** An LLM-powered news bot summarizes the trial as "Successful" because it saw the word "improvement," missing the "not statistically significant" caveat in the footnote.

3. **Market Reaction:** Sentiment-driven algos buy, pushing the stock up 5% in 30 seconds.

4. **Correction:** Human analysts (or more sophisticated "Deep-Reasoning" models) read the full report, realize the trial failed, and the stock crashes 10% an hour later.

5. **EMH Verdict:** The market was *temporarily* inefficient due to LLM-generated noise, before reverting to a semi-strong efficient state.

3. Mean Reversion and Efficiency

Mean Reversion is the statistical tendency for prices to return to their historical average. From an EMH perspective, mean reversion is an "anomaly." It suggests that the market occasionally overshoots (due to noise) and then corrects, providing a predictable window for profit.

Summary: Efficiency in the 2020s

| Feature | 1970s Market | 2020s Market |

| :--- | :--- | :--- |

| **Information Source** | Quarterly reports, Ticker tape | Real-time APIs, Twitter, GitHub |

| **Processing Agent** | Floor traders, Manual analysts | LLMs, HFT Algos, Sentiment Engines |

| **Efficiency Speed** | Days/Hours | Milliseconds |

| **Dominant Noise** | Rumors, Newsletters | Viral social media, AI-generated "Slop" |