💰 quantitative finance

AI-Driven Alpha Decay: Algorithmic Homogenization, Reflexive Signal Erosion, and the Paradox of Intelligent Markets

This paper argues that the widespread adoption of AI in investment strategies creates a self-defeating cycle of signal homogenization and accelerated decay, ultimately eroding excess returns, triggering extinction cascades, and increasing systemic fragility despite improved price discovery.

Original authors: Shuchen Meng, Xupeng Chen

Published 2026-05-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Shuchen Meng, Xupeng Chen

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Idea: The "Smart" Trap

Imagine a group of treasure hunters looking for gold in a vast field. In the beginning, only a few hunters have metal detectors. They find gold easily, and the field is full of riches.

Now, imagine that everyone suddenly buys the exact same high-tech metal detector, programmed with the exact same software, and trained on the exact same map.

At first, this seems great. But the paper argues that this creates a paradox: The more "smart" everyone gets, the less money anyone makes.

The authors call this the "Red Queen" trap (named after a character in Alice in Wonderland who has to run faster and faster just to stay in the same place). In the world of AI investing, everyone is running faster (using more powerful AI), but the "gold" (extra profit) is disappearing so quickly that no one actually gets richer.

The Three Ways AI Kills Its Own Profits

The paper explains that AI destroys its own profits through three specific mechanisms:

1. The "Crowded Room" Effect (Signal Crowding)

The Analogy: Imagine a crowded party where everyone is wearing the same outfit and listening to the same music. If one person says, "Look, there's a free pizza over there!" everyone rushes to the pizza at the exact same time. The pizza is gone in a split second, and no single person gets a big slice.
The Paper's Claim: AI systems are trained on the same data (stock prices, news, satellite images) and use similar models. When they spot a "profit opportunity," they all try to buy or sell at the same time. This happens so fast (in milliseconds) that the opportunity vanishes before any single investor can make a real profit.
The Result: The "half-life" of a profit opportunity has shrunk. Before AI, a good investment idea might last 5 to 7 years. Now, with high AI adoption, it lasts only about 18 months.

2. The "Self-Fulfilling Prophecy" Trap (Performative Erosion)

The Analogy: Imagine a weather forecaster who predicts rain. Because of the prediction, everyone buys umbrellas and stays inside. The streets stay dry. The next day, the forecaster looks at the data, sees it didn't rain, and thinks, "My prediction was wrong!" So, they change their model. But the reason it didn't rain was because of the prediction. The act of predicting changed the reality, making the prediction tool less accurate over time.
The Paper's Claim: When AI trades based on a signal, it changes the market prices. The AI then looks at the new prices to learn for next time. But the new prices are "contaminated" by its own previous trades. It's like trying to learn how to swim by looking at a pool you just drained. The signal gets "eroded" or degraded because the AI is constantly eating its own tail.

3. The "Arms Race" (Red Queen Competition)

The Analogy: Two runners are in a race. To win, Runner A buys better shoes. Runner B sees this and buys even better shoes. Runner A buys jetpacks. Runner B buys a rocket. They are spending millions on equipment, but they are still running at the same speed relative to each other. They are just exhausted and broke.
The Paper's Claim: Because profits are shrinking, investors feel forced to spend more on AI to stay ahead. They buy more data, more computers, and smarter models. But because everyone does this, the competition just gets fiercer, and the profits shrink even more. In the end, the industry spends a fortune on AI, but the net profit is zero. The only thing left is a fragile system that is prone to crashing.

The Four Key Findings

Profits Die Faster: The more AI is used, the faster investment ideas become useless. It's not a straight line; it gets worse faster and faster (convex decay).
The Domino Effect: When one type of investment strategy stops working (because too many AI are using it), the AI systems all rush to the next best strategy, killing that one too, and then the next. It's a cascade of "strategy deaths."
The Zero-Sum Game: In a market where everyone uses the same AI, the total profit from "smart" trading is zero. You are just paying for the privilege of running in place.
Efficiency vs. Safety: AI makes markets work better on a normal day (prices are accurate), but it makes them much more dangerous during a crisis. Because everyone reacts the same way to bad news, a small problem can turn into a massive crash (like the 2010 Flash Crash) because all the algorithms panic at once.

The Solution? Diversity is Key

The paper suggests that the only thing saving the market from total collapse is human diversity.

The Analogy: If a forest is made of only one type of tree, a single disease can kill the whole forest. But if the forest has many different types of trees, the disease might kill some, but the others survive.
The Paper's Claim: Human investors think differently. They don't all use the same data or the same models. This "messiness" actually protects the market. If everyone is an AI, the market is fragile. If there is a mix of AI and humans, the humans act as a shock absorber when the AI all panic together.

Summary

The paper concludes that AI is a powerful tool that has made markets smarter and faster. However, when everyone uses the exact same AI tools, the collective result is a loss of profit and an increase in danger. It is a case where individual intelligence leads to collective foolishness. The market is becoming "intelligent" but also incredibly fragile, like a house of cards built by a supercomputer.

Technical Summary: AI-Driven Alpha Decay

Problem Statement

The paper addresses a fundamental paradox in modern asset management: while Artificial Intelligence (AI) theoretically enhances individual price prediction accuracy, the mass adoption of similar AI systems appears to erode the aggregate excess returns (alpha) they seek to capture. Despite high adoption rates (91% of institutions as of 2024) and algorithmic trading dominating 60–80% of US equity volume, aggregate hedge fund performance has declined, with average funds underperforming simple 60/40 portfolios since 2009. The central research question is whether AI-driven investing contains the seeds of its own destruction by endogenously destroying the market inefficiencies required for profitable trading.

Methodology

The authors construct a three-layer theoretical equilibrium model integrating market microstructure, machine learning dynamics, and game theory, validated through calibrated simulations against real-world data.

Theoretical Framework

The model is built upon three interacting layers:

Layer 1: Signal Crowding and Alpha Dynamics.
- Signal Structure: Tradeable signals ( $f_k$ ) follow Ornstein–Uhlenbeck processes with a natural mean-reversion rate ( $\theta$ ).
- AI Homogeneity: AI investors observe signals with a correlation parameter ( $\rho$ ) reflecting shared training data and model architectures. Human investors observe independent signals.
- Mechanism: As the AI adoption rate ( $\phi$ ) increases, correlated order flows from $N\phi$ agents accelerate the arbitrage of mispricings. This introduces an endogenous decay component $\delta(\phi) = N\phi\rho a / \lambda(\phi)$ , where $\lambda$ is the price impact.
- Result: The effective decay rate becomes $\theta_{eff} = \theta + \delta(\phi)$ , shortening the lifespan of tradeable signals.
Layer 2: Reflexive Signal Erosion (Performative Prediction).
- Mechanism: AI systems retrain on data contaminated by the price consequences of their own prior trading. This creates a feedback loop where the "observed" return is the "counterfactual" return minus the price impact of the aggregate AI trade.
- Micro-foundation: Using a performative prediction framework (Perdomo et al., 2020), the authors derive that the estimated signal coefficient attenuates as $\hat{\beta} = \beta^{CF} \cdot \frac{\theta}{\theta + \delta(\phi)}$ .
- Dynamics: This leads to a "Signal Extinction Cascade." When adoption exceeds a critical threshold $\phi^*$ , the erosion of signal quality outpaces the regeneration of new mispricings, causing specific factor classes to sequentially lose all predictive power.
Layer 3: Red Queen Equilibrium.
- Game Structure: Institutions choose between AI and human strategies based on gross alpha, investment costs, and career concerns (reputational costs of deviating from the industry consensus).
- Equilibrium: A "Red Queen" equilibrium emerges where institutions overinvest in AI to avoid competitive obsolescence, even though the net alpha from AI investment is driven to zero. The primary output of this arms race is not returns, but systemic externalities (correlated tail risk).

Empirical Validation

The authors validate the model using three approaches:

Portfolio Convergence: Calibrated simulations matched to 99.5 million SEC Form 13F holdings (2013–2024) show a 42% increase in institutional portfolio convergence (measured by pairwise cosine similarity), with a 58% increase specifically among AI-adopting funds. Structural breaks align with AI technology adoption waves (e.g., GPT-4 release).
Return Dispersion: Simulations of hedge fund returns (2010–2024) demonstrate a 29% decline in cross-sectional return dispersion for AI funds compared to a 10% decline for human-managed funds.
Fragility Analysis: The model simulates the 2010 Flash Crash, estimating a "reflexive multiplier" ( $M \approx 1.4$ ) where algorithmic homogeneity amplified a fundamental shock by approximately 39%, consistent with observed market dynamics.

Key Contributions and Results

1. The Alpha Half-Life Theorem

The paper derives the alpha half-life $h(\phi)$ as:
$h(\phi) = \frac{\ln 2}{\theta + \delta(\phi)}$
where $\delta(\phi)$ is convex-increasing in adoption.

Finding: The half-life of tradeable signals is convex-decreasing in AI adoption.
Quantification: At current adoption levels ( $\phi \approx 0.7, \rho \approx 0.6$ ), the model implies a signal half-life of 18 months, a sharp contraction from the pre-AI era estimate of 5–7 years.

2. Signal Extinction Cascade

The model establishes that alpha decay is not uniform but proceeds through discrete "extinction waves."

Mechanism: As one signal class (e.g., momentum) is arbitraged away due to high crowding, AI capital reallocates to remaining signals, accelerating their decay.
Prediction: Factors with high crowding intensity and low regeneration rates (e.g., momentum) die first, followed sequentially by others.

3. The Red Queen Impossibility

In the monoculture equilibrium ( $\phi \to 1$ ), the net alpha from AI investment is identically zero ( $\alpha_{net} = 0$ ).

Paradox: While individual AI systems extract genuine informational value, the aggregate extraction arbitrages away the very patterns being exploited.
Overinvestment: The privately optimal level of AI investment exceeds the socially optimal level due to unpriced externalities (signal crowding and systemic fragility).

4. Efficiency–Fragility Tradeoff

The paper identifies a fundamental tradeoff in market welfare:

Efficiency: AI adoption improves average price discovery (allocative efficiency).
Fragility: It increases tail risk and susceptibility to flash crashes.
Result: The adoption level that maximizes price discovery is strictly higher than the level that minimizes systemic fragility. The market equilibrium overshoots even the efficiency optimum, leading to a "fragile efficiency."

Significance and Claims

The paper claims to provide the first formal equilibrium model that jointly determines the endogenous rate of alpha decay, the feedback from trading to signal quality, and the equilibrium level of AI investment.

Theoretical Distinction: Unlike existing literature focusing on systemic risk or standard performative prediction, this work formalizes performative signal erosion mediated by market microstructure (price impact and order flow) and introduces the Red Queen equilibrium as an explanation for declining hedge fund performance despite rising AI investment.
Empirical Relevance: The model offers a unified explanation for the "quant earthquake" of 2007, the 2010 Flash Crash, and the recent compression of factor premiums, attributing them to the interaction of signal crowding and reflexive feedback.
Policy Implication: The authors argue that the "competence illusion"—where managers believe AI is superior because signals appear strong in early adoption phases—leads to a monoculture trap. They suggest that human diversity acts as a resilience mechanism, providing a "diversity premium" that reduces tail risk when AI strategies fail collectively.

The paper concludes that while AI improves individual price predictions, collective adoption creates a qualitatively different market environment where the pursuit of alpha becomes a self-defeating arms race, resulting in a market that is highly efficient on average but structurally fragile.