Eighteen Months: The Shrinking Half-Life of Profitable Trading Signals and Citadel's AI Lab Talent Raid

A Number That Converts to Career Length

An effective researcher's career spans about thirty years. A money-making trading signal, according to a model from two NYU researchers, now loses half its excess returns in roughly eighteen months. Before AI became widespread, that figure was five to seven years.

On the same day, Bloomberg-affiliated media reported that Citadel is expanding its quant team—and reaching into AI labs for hires. Reading these two stories together is far clearer than reading them apart.

What Citadel Is Doing

Navneet Arora, Citadel's global head of quant strategies, offered the numbers: a team of roughly 180 people including financial researchers and data engineers, planning double-digit headcount growth over the coming year.

A newly formed team is led by Alexey Poyarkov, who joined recently from TGS Management, focused on global equities systematic trading. Beyond traditional university recruiting, a new channel has appeared: AI labs like Google DeepMind. COO Gerald Beeson said this puts them in direct competition with OpenAI and Anthropic for some hires.

Arora gave a distinctive answer on whether AI will replace researchers. AI hasn't replaced human researchers, he said—it has lowered the importance of coding ability while raising the value of a different type of person: those who can generate investment ideas and guide machines to turn ideas into profitable trades. In his words, one person can now effectively manage two, four, or in some cases even more processes; humans are becoming managers of machines, and such people are not as easily replaced as machines.

The performance context: Citadel's Tactical Trading fund, combining fundamental equity investing with quant strategies, is up 24.7% this year through August, with roughly 20% annualized returns since its 2008 inception. Citadel manages about $76 billion in assets and currently allocates billions to quant strategies.

How the Eighteen Months Was Calculated

The widely cited paper is by NYU's Shuchen Meng and Xupeng Chen: *AI-Driven Alpha Decay: Algorithmic Homogenization, Reflexive Signal Erosion, and the Paradox of Intelligent Markets*, arXiv:2605.23905, SSRN version dated February 28, 2026.

The core is a half-life formula: \(h(\phi) = \ln 2 / [\theta + \delta(\phi)]\). Here \(\theta\) is the signal's natural mean-reversion rate, and \(\delta(\phi) = N\phi\rho a / \lambda(\phi)\) depends on AI adoption rate \(\phi\), algorithmic signal correlation \(\rho\) (reflecting shared training data and model architectures), and endogenous price impact \(\lambda(\phi)\).

Using current adoption \(\phi \approx 0.7\) and \(\rho \approx 0.6\), the researchers compute an 18-month half-life, versus 5–7 years pre-AI.

Three channels reinforce each other. Signal crowding means everyone buys the same stocks. Reflexive erosion means the act of trading itself erases the market inefficiencies that generate alpha. Red Queen competition means everyone accelerates and no one pulls ahead.

The paper presents four theoretical results:

• Half-life theorem: signal lifespan declines convexly with AI adoption—each additional AI participant shortens exploitable patterns' lifespan at an accelerating rate.
• Signal extinction cascade: past a critical threshold, the decay of one signal class triggers accelerated competition for the rest.
• Red Queen impossibility: in a single-population equilibrium, net alpha is zero no matter how heavily everyone invests in AI.
• Fragility-efficiency trade-off: the adoption level that maximizes price discovery is strictly higher than the level that minimizes systemic fragility.

The empirical section calibrates portfolio convergence using SEC 13F filings—99.5 million position records from 2013 to 2024—simulating a 42% rise in institutional portfolio convergence over the sample period.

The authors hedge their language: each new AI participant shortens all exploitable patterns' lifespan at a faster rate, and when everyone uses similar AI, the aggregate outcome is qualitatively different from the simple sum of individual gains.

Two Other Research Lines From the Same Period

The University of Liechtenstein's Advije Rizvani, Giovanni Apruzzese, and Pavel Laskov built 10 LLM-based trading models using sentiment analysis to predict a set of stock prices. Over a 14-month investment period ending April 2025, all posted positive returns. Then researchers made changes to financial news headlines nearly imperceptible to human readers—swapping letters for visually similar characters or embedding hidden text. Every model was misled. In the worst case, manipulating a single stock on a single trading day cut one model's overall returns by about 18 percentage points.

A third line comes from Jerry Bell, Victor Haghani, and James White of Elm Partners Management. They had four popular AI models read advance editions of Wall Street Journal front pages, then bet on the S&P 500 and US Treasuries. Claude and ChatGPT's directional accuracy matched top macro traders, exceeding 50%. The experiment also exposed a key blind spot related to risk-taking.

How Much to Believe

Four boundaries:

1. The half-life paper is primarily theoretical modeling; the authors themselves say empirical validation is still in progress. The 42% convergence figure comes from simulated calibration against 13F positions, not directly observed alpha decay. 2. The 18-month figure rests on two parameter values (adoption ~0.7, signal correlation ~0.6). How these were estimated, and whether different values would flip the conclusion, lacks sufficient sensitivity analysis in the paper. 3. Crowded trades were a feature of US markets before AI, especially in tech stocks, without catastrophic consequences. This objection was raised explicitly in coverage. 4. The Liechtenstein experiments were controlled: 10 models, one set of stocks, 14 months, artificially designed perturbations. They prove an attack surface exists, not that it is exploited at scale in reality.

What's Worth Watching

For AI quant practitioners, this material shifts the question from "can AI help me make money" to "how long before my profits get arbitraged away." Citadel's expansion and AI lab recruiting happen against exactly this backdrop. Arora's "one person managing four processes" framing points to the other side of the same coin: once tools compress execution costs, what becomes scarce is the idea itself and the speed of validating it.

If the half-life drop from five-to-seven years to eighteen months holds, the biggest impact is on strategy R&D scheduling logic, not model selection.

Watch List

• Whether Meng and Chen's empirical section gets a formal release, and whether the 18 months gains direct observational support.
• Parameter sensitivity: the half-life range under different adoption and correlation values.
• Citadel's expansion results: team size and Tactical Trading performance a year out.
• Real-world scale of the attack surface: how often headline-style manipulations appear in live markets.
• The risk-taking blind spot: Elm's full report, if published, deserves a standalone read.

Sources

1. Bloomberg-syndicated report on Citadel's quant hiring, via Sina Finance, 2026-09-25 2. Shuchen Meng, Xupeng Chen, *AI-Driven Alpha Decay: Algorithmic Homogenization, Reflexive Signal Erosion, and the Paradox of Intelligent Markets*, arXiv:2605.23905 / SSRN 6349698, 2026-02-28 3. "AI arms race on Wall Street raises new concerns over crowded trades," The Straits Times via Sina Finance, 2026-07-01 4. University of Liechtenstein LLM robustness experiments (Rizvani, Apruzzese, Laskov), as relayed in above coverage 5. Elm Partners Management AI directional-accuracy simulated trading experiment, as relayed in above coverage