A generative-AI 'copilot' speeds equity research and sharpens qualitative signals in backtests on A-share and U.S. large-cap stocks, but model hallucinations, prompt sensitivity and data‑leakage risks can distort trading unless tightly governed; traceable prompts, cross‑source checks and model risk controls are essential for safe deployment.

Main Finding

Deepseek (a generative-AI copilot) can meaningfully improve equity-research throughput and produce incremental, tradable signals when its text outputs are converted into structured factors and embedded in disciplined portfolio rules. The largest practical benefits are faster event understanding, timelier qualitative signals around disclosures, and scalable contradiction/uncertainty detection. However, gains are conditional: model hallucination, prompt sensitivity, timestamp/data-leakage risks and higher turnover can erase benefits unless strong governance, audit logging, and cross-source verification are enforced. Deepseek is therefore most valuable as a governed research-assist tool—not an autonomous trading agent.

Key Points

• Role and niche

  • GAI adds value primarily at the signal-construction and research-efficiency layers (summarization, classification, contradiction detection, confidence tagging).
  • It is most helpful in event-intense settings (earnings season, policy shocks) and large universes where manual reading capacity is the binding constraint.
  • Not reliable when asked for unconstrained buy/sell recommendations.

• Workflow and governance

  • Best practice is a hybrid workflow: Deepseek produces structured textual outputs that are mapped into numeric signals and fed into conventional factor-based portfolio construction.
  • Critical governance elements: standardized prompt templates, immutable logging (prompt ID, model version, document hash, timestamp), two-model or deterministic consistency checks, analyst sign-off for high-impact signals, and embargo/timestamp discipline to prevent leakage.

• Behavior of GAI-derived signals

  • Textual signals decay faster (shorter half-life) than valuation factors; they are most informative near event timestamps.
  • Signal distribution is heavy-tailed: a minority of outputs contain useful incremental information; many outputs are noisy or hallucinated.
  • Quality filters (confidence thresholds, cross-source weighting) improve signal-to-noise and reduce turnover.

• Performance patterns (case study)

  • Compared three settings: Baseline (traditional factors), GAI-Naive (add unconstrained Deepseek scores), GAI-Governed (Deepseek under verification).
  • GAI-Naive produced occasional outsized gains but higher drawdowns and turnover from spurious signals.
  • GAI-Governed delivered more stable improvements in risk-adjusted metrics vs. Baseline by removing low-confidence signals and reducing churn.
  • Combining GAI signals with classic factors yielded better diversification than GAI-only implementations.

Data & Methods

• Universe & timeframe

  • Representative large-cap equities (A-share and U.S. large-cap), rolling monthly implementation spanning multiple market regimes (tightening/easing cycles, high/low volatility).

• Inputs

  • Structured: daily price-volume, fundamentals.
  • Unstructured: earnings-call transcripts, policy news, firm announcements; texts aligned to first-public timestamps.

• Deepseek usage

  • Tasks: event summarization, sentiment polarity classification, risk-factor extraction, uncertainty/confidence tagging, contradiction detection, analyst-style narratives.
  • Prompt protocol: standardized templates (e.g., earnings-call outputs must include key positives, key negatives, guidance direction, confidence); automated re-queries for formatting failures.
  • Audit trail: model version, timestamp, prompt template ID, document hash; two-model consistency checks for critical events.

• Signal engineering and portfolio construction

  • Convert textual outputs into standardized numeric variables (SGAI_i,t: event-sentiment, narrative-uncertainty, governance-risk).
  • Aggregated score used in ranking: Score_i,t = α1 SGAI_i,t + α2 Mom_i,t + α3 Val_i,t − α4 Vol_i,t. Coefficients frozen after training.
  • Portfolio: benchmark-neutral long-short deciles and long-only top-quintile tests; sector and single-name caps, turnover limits; conservative transaction-cost assumptions.

• Experimental comparisons & diagnostics

  • Three settings compared: Baseline, GAI-Naive, GAI-Governed.
  • Metrics: annualized return, Sharpe ratio, max drawdown, turnover-adjusted information ratio, earnings-event hit rate, signal half-life, performance under stress.
  • Robustness checks: regime sub-samples, alternative signal aggregations (median, source-weighted), rebalancing frequencies, transaction-cost/slippage stress, model-version backtests, cross-universe transfer.

Implications for AI Economics

• Market microstructure and efficiency

  • GAI can reduce informational latency by accelerating narrative processing, which compresses informational frictions for institutions that adopt it—potentially narrowing transient mispricings around events.
  • Adoption may change the value of “speed vs. breadth” in research: faster narrative processing becomes a valuable, monetizable service.

• Alpha production and persistence

  • Text-derived alpha is conditional and ephemeral (short half-life). Economic value accrues where bounded, high-quality textual signals change marginal allocation decisions (names near cutoffs), rather than confirming consensus.
  • Governance acts as a selection mechanism (filter) that converts noisy GAI outputs into useful signals; thus model-risk controls are integral to realizing economic benefits—not optional overhead.

• Organizational and investment implications

  • Firms need to invest in infrastructure: prompt/version registries, immutable document logs, verification pipelines, role-based responsibility matrices (research, risk, compliance, tech).
  • Human-AI collaboration models (GAI as copilot) preserve accountability and are practically superior to end-to-end automation for equity selection given current hallucination and instability risks.

• Policy and regulatory considerations

  • Auditability and reproducibility (prompt/version logs, timestamp discipline) will be important for compliance and model-risk management; regulators and internal risk teams should treat GAI-driven pipelines like other quantitative models.
  • Data-leakage and timestamp integrity are non-trivial systemic risks for text-driven strategies and require stricter controls than many numeric-only pipelines.

• Broader economic effects

  • Widespread, governed GAI adoption could increase overall informational efficiency but also raise competitive pressures (higher churn, shorter signal horizons) and operational costs (governance, auditing).
  • The net social value depends on the balance of improved information processing versus amplification of noise/overreaction when governance is weak.

Summary takeaway: Deepseek-style GAI can be a high-value productivity and signal-enhancement tool in equity investing if used within auditable, versioned, and human-supervised processes. Without those risk controls, its prompt sensitivity, hallucination propensity, and leakage risks can negate or reverse economic benefits.