AI Premium

Battle Count: 50

Highly relevant to quantitative trading. The paper directly constructs a tradable AI factor and demonstrates a statistically significant AI premium (64.1 bps/week value-weighted, ~56 bps after FF5+Mom adjustment). The long-short AI-beta portfolio strategy is explicitly characterized. The decomposition into salient components (closed-source vs open-weight, paid/core vs new users, seasoned vs non-seasoned, long vs short prompts) provides actionable sub-factors for factor investing. The international evidence (17.9 bps in developed markets, absent in emerging/China) informs cross-market strategies. The event study around model releases provides timing signals. The occupation/skill mapping offers fundamental insights for sector rotation. The agentic AI factor, while imprecise, represents an emerging alpha source.

The asset pricing methodology (rolling regressions, portfolio sorts, Fama-MacBeth) is standard in quantitative finance. However, implementation requires: (1) access to proprietary OpenRouter licensed data at user-model-day granularity, (2) construction of multiple AI sub-factors from decomposed consumption data, (3) mapping firm exposures to occupations via BLS weights and O*NET ratings, (4) hand-collected AI model release calendar for event studies, (5) international data from Compustat Global and Datastream China. The core portfolio construction and regression analysis is moderate complexity, but the data pipeline for AI consumption factor construction is non-trivial and data-dependent.

The core dataset (OpenRouter licensed proprietary data covering 380 trillion tokens) is not publicly available. The analysis requires access to anonymized user-model-day level AI consumption data from OpenRouter, Inc. Standard financial data (CRSP, Compustat, Fama-French factors) are available through academic subscriptions, but the primary AI consumption data is proprietary and licensed. The paper states data is used under license and original data are not redistributed or disclosed.