Expanding short-selling access in China’s stock market nudges firms toward AI: companies exposed to the short-selling mechanism raise AI-related investment and patenting, partly driven by increased R&D spending.

Main Finding

The paper uses a quasi-natural experiment from the phased expansion of China’s margin trading and securities‑lending (short‑selling) program (A‑share universe, 2015–2025) and finds that inclusion in the short‑selling/margin‑trading target list significantly increases firms’ measured AI adoption. The effect is partially mediated by higher R&D investment. Results are robust to using AI‑related patent counts as an alternative dependent variable.

Key Points

• Causal design: multi‑period difference‑in‑differences (Treat × Post) exploiting exogenous, regulator‑driven staged inclusion of firms in the margin/short‑selling list.
• Main empirical result: Treat_Post coefficient ≈ 0.0813 (p < 0.01) on AI_Index (text‑based measure). Given AI_Index mean ≈ 0.873, this is roughly a 9.3% increase relative to the sample mean.
• Mediation: Treat_Post increases firm R&D (coefficient ≈ 0.0062, p < 0.01; R&D mean ≈ 0.046 → ≈13.5% relative increase). When R&D is added, the Treat_Post coefficient on AI_Index falls from 0.0813 to 0.0576 but remains statistically significant, indicating a partial mediation via R&D.
• Robustness: Using ln(1 + AI_patent_count) as the dependent variable, Treat_Post ≈ 0.0691 (p < 0.01), supporting the baseline finding.
• Proposed channels: (1) improved information environment / price discovery and (2) stronger innovation incentives prompting greater R&D and innovation output.
• Sample & scale: Panel of A‑share listed firms, 2015–2025, N = 18,462 firm‑year observations; ~24% of observations are post‑treatment (Treat_Post mean ≈ 0.238).

Data & Methods

• Data sources: CSMAR margin trading database (treatment timing and eligible securities), CSMAR Annual Report Full‑text Database and Juchao for text mining, National IP Administration patent database, and CSMAR financial/governance modules.
• Dependent variables:
  • AI_Index = ln(1 + frequency of AI‑keywords in annual reports) — primary measure of firm AI application emphasis.
  • Robustness: AI_patent = ln(1 + number of AI‑related patents in year).
• Independent variable: Treat × Post, where Treat = 1 for firms included in the margin/short‑selling target list and Post = 1 for years after inclusion.
• Mediator: R&D investment (absolute or as share of assets); supplementary mediator: number of invention/utility patents.
• Controls and estimation: firm characteristics (size, leverage, ROE, growth, ownership type, etc.), firm fixed effects, year fixed effects, standard errors clustered at firm level.
• Identification checks: authors report event‑study/parallel‑trend testing and multiple robustness checks (alternative dependent variable, etc.).
• Key reported statistics: baseline R‑squared ≈ 0.187; standard errors clustered at firm level.

Implications for AI Economics

• Market institutions matter for technology adoption: Capital‑market regulatory choices (here, expanding short‑selling eligibility) can alter firms’ incentives and external monitoring in ways that accelerate corporate AI uptake.
• Two levers for policy and investors:
  • Regulators: carefully designed, transparent market access policies (e.g., margin/short‑selling eligibility) can be used as indirect tools to strengthen governance and encourage productive risky investments like AI.
  • Investors & analysts: short‑selling channels that improve price discovery and scrutiny may increase managerial incentives to invest in long‑term, uncertain technologies; investor engagement strategies should account for these dynamics.
• Mechanism insight: an external governance channel works partly by increasing R&D spending—suggesting that policies that tighten monitoring/expectations can convert into greater innovation effort, which lowers barriers to AI adoption.
• Research and evaluation: when assessing determinants of AI diffusion, include capital‑market institutional features (beyond firm internal factors) as potentially important drivers.

Limitations & Notes for Future Research

• Measurement: the primary AI measure (annual‑report keyword frequency) captures signaling/strategic emphasis as well as substantive adoption; patents partially mitigate this but have their own lags and industry biases.
• External validity: results derive from China’s institutional rollout; transferability to other markets with different regulatory frameworks requires caution.
• Remaining endogeneity risks: authors argue eligibility is regulator‑determined (exogenous), but unobserved time‑varying selection on eligibility criteria could remain; event‑study and fixed effects mitigate but do not wholly eliminate concerns.
• Outcomes: the study documents increased AI adoption signals and R&D but does not (in this paper) quantify downstream productivity or profitability effects of increased AI uptake—an avenue for follow‑up work.

Overall, the paper provides empirical evidence that capital‑market governance mechanisms—here, the short‑selling/margin trading regime—can materially influence firm decisions to adopt AI, partly by stimulating R&D investment.