Autonomous Market Intelligence: Agentic AI Nowcasting Predicts Stock Returns

TL;DR

The paper tests whether fully agentic AI can nowcast stock returns in a strict out-of-sample framework by having an autonomous LLM search the live web to rate Russell 1000 stocks. It constructs a unique, irreproducible dataset of AI-attractiveness scores and daily predictions, then forms implementable Top-20 portfolios using open-to-open returns and evaluates them with a Fama-French six-factor model. The main findings show robust daily alphas (e.g., $ ext{alpha}_{daily} = 0.184 ext{%}$) and a Sharpe ratio of $2.43$ for Top-20 portfolios, driven by a growth, low-volatility tilt and concentrated in the very top ranks; bottom-ranked stocks do not produce reliable negative alphas. The results imply that agentic AI processes real-time information in a qualitatively different way from prior methods, offering actionable, high-conviction opportunities even after realistic transaction costs, and establish a living benchmark for AI evaluation in finance. These contributions have broad implications for AI design, market efficiency, and how we assess cognitive capabilities of autonomous systems in dynamic, adversarial settings.

Abstract

Can fully agentic AI nowcast stock returns? We deploy a state-of-the-art Large Language Model to evaluate the attractiveness of each Russell 1000 stock daily, starting from April 2025 when AI web interfaces enabled real-time search. Our data contribution is unique along three dimensions. First, the nowcasting framework is completely out-of-sample and free of look-ahead bias by construction: predictions are collected at the current edge of time, ensuring the AI has no knowledge of future outcomes. Second, this temporal design is irreproducible -- once the information environment passes, it can never be recreated. Third, our framework is 100% agentic: we do not feed the model news, disclosures, or curated text; it autonomously searches the web, filters sources, and synthesises information into quantitative predictions. We find that AI possesses genuine stock selection ability, but only for identifying top winners. Longing the 20 highest-ranked stocks generates a daily Fama-French five-factor plus momentum alpha of 18.4 basis points and an annualised Sharpe ratio of 2.43. Critically, these returns derive from an implementable strategy trading highly liquid Russell 1000 constituents, with transaction costs representing less than 10\% of gross alpha. However, this predictability is highly concentrated: expanding beyond the top tier rapidly dilutes alpha, and bottom-ranked stocks exhibit returns statistically indistinguishable from the market. We hypothesise that this asymmetry reflects online information structure: genuinely positive news generates coherent signals, while negative news is contaminated by strategic corporate obfuscation and social media noise.

Figures (7)

• Figure 1: Cumulative Returns of AI-Selected Portfolio vs. Benchmark. This figure plots the cumulative value-weighted returns (base 100) of the Top-20 stocks selected daily by the AI model's attractiveness score (solid orange line) against the Russell 1000 index (dashed gray line). The AI-selected portfolio accumulates approximately 50% returns over the sample period compared to 26% for the benchmark. The performance wedge widens consistently over time, suggesting a persistent information advantage rather than exposure to isolated extreme events. Sample period: April 2025 through January 2026.
• Figure 2: Cumulative returns of the Top-20 portfolios under different signal horizons. This figure plots the value-weighted cumulative returns of portfolios of Top-20 stocks ranked by AI's daily, weekly, monthly, and quarterly attractiveness scores. Each panel shows the cumulative return (base 100) for the Top-20 portfolio and the Russell 1000 benchmark over the sample period.
• Figure 3: Top versus Bottom Alpha Performance. This figure reports daily Fama-French six-factor alphas (%) for value-weighted portfolios of the top $N$ (blue) and bottom $N$ (orange) stocks ranked by AI attractiveness. Panels correspond to daily, weekly, monthly, and quarterly signal horizons. Error bars indicate 90% confidence intervals using Newey-West standard errors (5 lags). Sample period: April 2025 onwards.
• Figure 4: Portfolio Turnover Ratios (Top-20). This figure plots the time series of portfolio turnover for the Top-20 stocks selected by the AI model. Panels correspond to daily, weekly, monthly, and quarterly signal horizons. Turnover is calculated as the ratio of new stocks entering the portfolio relative to the total number of holdings. The dashed orange line indicates the sample average. Periods of zero turnover reflect operational data availability constraints rather than signal persistence. Sample period: April 2025 onwards.
• Figure 5: Granular Cross-Sectional Alpha Distribution. This figure plots daily Fama-French six-factor alphas (%) for 50 rank-sorted value-weighted portfolios (approximately 20 stocks each). The $x$-axis represents rank group midpoints; lower values indicate stocks ranked more attractive by the AI signal. The dashed red line indicates zero alpha. Error bars represent 90% confidence intervals using Newey-West standard errors (5 lags). Sample period: April 2025 onwards.