IP Risks In Machine-Learning Models Predicting Polish Stock Market Patterns.
1. Copyright Risks in Training Financial ML Models
Core Issue
ML models are trained on historical stock data, financial reports, analyst commentary, and news feeds. While raw market prices may not always be protected, compiled datasets, reports, and structured databases often are.
Key Risk Areas
- Unauthorized scraping of financial databases
- Use of proprietary analyst reports
- Copying structured datasets protected as compilations
Case Law 1: Feist Publications v. Rural Telephone Service
Principle
- Facts are not copyrightable, but original selection/arrangement is protected.
Application
- Stock prices themselves = facts (not protected)
- But:
- Curated datasets (e.g., sector-wise classification, weighted indices)
- Structured historical trend databases
👉 ML models trained on such curated datasets may infringe copyright if copied without permission.
Case Law 2: British Horseracing Board Ltd v William Hill Organization Ltd
Principle
- EU recognizes sui generis database rights protecting substantial investment in databases.
Application in Poland (EU Law Applies)
- Financial datasets (e.g., Warsaw Stock Exchange structured data)
- If ML developers extract substantial portions, it may violate:
- Database rights (even if no copyright infringement)
👉 This is one of the biggest risks in EU-based ML finance systems.
2. Text and Data Mining (TDM) Exceptions vs Commercial Use
EU law allows limited Text and Data Mining (TDM) under directives like DSM Directive.
Risk
- TDM exceptions apply only:
- For research or
- If rights holders have not opted out
👉 Commercial ML trading systems often fall outside safe exceptions.
Case Law 3: Infopaq International A/S v Danske Dagblades Forening
Principle
- Even small extracts (11 words) can be protected if they reflect originality.
Application
- ML models ingest:
- News headlines
- Financial summaries
👉 Even partial ingestion of proprietary text data can trigger copyright infringement.
3. Output Liability: Are Predictions Infringing?
ML models generate:
- Trading signals
- Pattern predictions
- Market summaries
Risk
- If outputs reproduce:
- Proprietary analysis
- Unique data structures
Case Law 4: SAS Institute Inc v World Programming Ltd
Principle
- Functionality and ideas are not protected, only expression.
Application
- ML model predictions:
- If they replicate function → OK
- If they replicate exact expression or structure → infringement
👉 Example: replicating proprietary trading signals format may create risk.
4. Trade Secret Risks in Financial ML
Core Issue
Many stock prediction models rely on:
- Proprietary trading algorithms
- Confidential datasets
- Institutional strategies
Case Law 5: Saltman Engineering Co Ltd v Campbell Engineering Co Ltd
Principle
- Confidential information must not be misused even if not patented.
Application
- Employees or contractors building ML models:
- Using confidential trading strategies
- Training models on proprietary hedge fund data
👉 Leads to trade secret misappropriation claims.
Case Law 6: PepsiCo Inc v Redmond
Principle
- Employees may be restricted if they are likely to disclose trade secrets.
Application
- Data scientists moving between firms:
- Risk of “leaking” algorithmic trading strategies into new ML models
👉 Critical for fintech startups in Poland hiring experienced quants.
5. Software and Algorithm Patent Risks
Issue
ML-based trading systems may involve:
- Predictive algorithms
- Automated trading execution systems
Case Law 7: Alice Corp v CLS Bank International
Principle
- Abstract ideas implemented on computers are not patentable unless inventive.
Application
- Stock prediction ML models:
- Pure mathematical models → not patentable
- Technical improvements → possibly patentable
👉 Risk:
- Patent invalidation
- Or infringement of existing fintech patents
6. Copyright in AI Models Themselves
Issue
Who owns:
- Trained ML models?
- Model weights?
Case Law 8: Navitaire Inc v EasyJet Airline Co Ltd
Principle
- Functionality is not protected, only code/expression.
Application
- ML models:
- Architecture (idea) → not protected
- Code + trained weights → may be protected
👉 Competitors replicating model logic may avoid infringement.
7. Data Ownership & Financial Market Regulation Overlap
Issue in Poland
Stock data may be governed by:
- Exchange licensing agreements
- Financial regulations (e.g., MiFID II in EU)
👉 ML developers may face:
- Contractual liability
- Regulatory penalties
- IP infringement simultaneously
8. AI Training Using News & Reports
Case Law 9: Authors Guild v Google Inc
Principle
- Large-scale data use may be fair use if transformative (US context).
EU Contrast
- EU (including Poland) is stricter:
- No broad “fair use” doctrine
👉 ML models trained on financial news:
- Safer in US
- Riskier in EU without licenses
9. Reverse Engineering Risks
Issue
Competitors may:
- Analyze model outputs
- Reconstruct strategy
Case Law 10: Bayer AG v Housey Pharmaceuticals Inc
Principle
- Use of tools for research may still infringe patents.
Application
- Reverse engineering ML outputs:
- May infringe patents or violate IP protections
Key IP Risks Summary
1. Copyright Risks
- Use of proprietary datasets
- News, reports, structured data
2. Database Rights (Critical in EU/Poland)
- Extraction of substantial financial datasets
3. Trade Secret Risks
- Misuse of confidential trading algorithms
4. Patent Risks
- Algorithm patent infringement or invalidity
5. Output-Based Liability
- Replicating proprietary insights
6. Contractual & Regulatory Risks
- Exchange licensing violations
- Financial compliance overlaps
Conclusion
ML models predicting Polish stock market patterns operate in a high-risk IP environment, particularly due to:
- Strong EU database rights
- Limited TDM exceptions
- Overlap between IP law and financial regulation
The biggest legal exposure lies in:
- Training data acquisition
- Use of proprietary financial datasets
- Misuse of confidential trading strategies
RELATED Blog
comments