When AI Delivers Investment Signals: Where Does Model Risk Begin in Portfolio Management?
AI has moved well beyond being merely a research topic in asset management. Models can analyse company and market data and derive signals for security selection. This raises a new question for portfolio managers: how does risk management change when a model becomes part of the investment process?
A recent example shows how tangible this development has become. On 10 September 2026, Pictet AI Enhanced Equity UCITS ETFs were listed on SIX SIX Swiss Exchange. According to SIX, a proprietary AI model identifies patterns in company and market data and translates them into signals for security selection, while the quantitative investment team oversees the process and portfolio construction.
The industry is also focusing increasingly on this topic. The AMAS Risk Management Day on 30 September 2026 is dedicated entirely to Artificial Intelligence and addresses topics including AI in Risk Management, Machine Learning, Governance and the question of whether AI itself can become an operational risk for a financial institution.
The Model Delivers a Signal – Responsibility Remains
The use of AI does not automatically change responsibilities within the investment process. What matters, therefore, is how strongly the model is integrated into decision-making: Does it provide research signals? Does it influence portfolio weights? Or are certain decisions partly automated?
FINMA identifies operational risks and model risks associated with AI, including insufficient robustness, correctness, stability, explainability and bias. It also refers to data-related risks, IT and cyber risks, dependencies on third parties, as well as legal and reputational risks.
For portfolio managers, the quality of the output is therefore not the only relevant issue. Another important question is: Do we understand when and why the model may no longer behave as expected?
Data Quality May Be More Important Than the Model Itself
A good model using poor data remains a poor investment tool. FINMA expressly notes that data may be incorrect, incomplete, outdated or unrepresentative.
Historical data may also contain patterns or biases that are carried forward into future predictions. Where third-party solutions are used, transparency regarding data sources and methodologies may also be limited.
This is particularly relevant in portfolio management when market regimes change. A model that identifies stable relationships in historical data does not necessarily retain those relationships under new market conditions.
A Backtest Is Not Ongoing Monitoring
FINMA also discusses testing and ongoing monitoring as important elements in managing AI applications. Examples include backtesting, out-of-sample testing, sensitivity analysis, stress testing and comparison with simpler benchmark models.
Monitoring changes in input data is also relevant in order to identify potential data drift.
For an investment process, this means: A model should not only perform convincingly at go-live. What also matters is whether its results remain plausible and stable under changing market and data conditions.
Explainability: Does Every Model Need to Be Fully Understood?
Not necessarily every technical detail. However, the more important a model becomes for decision-making, the more important it is to be able to critically assess its outputs.
FINMA notes that results from AI applications may in some cases be difficult to understand, explain or reproduce. Where decisions need to be explained to investors, clients, supervisory authorities or auditors, FINMA has examined explainability in greater depth.
This includes understanding the relevant drivers of the model and its behaviour under different conditions.
For portfolio managers, this leads to an important point: A good signal alone does not constitute adequate governance.
What Does This Mean for Risk Management?
FINMA does not prescribe a specific checklist for AI in portfolio management. Its supervisory communication instead describes risks and measures that it has observed or reviewed in the course of its supervisory activities.
These include, among other things, an inventory of relevant AI applications, risk classification, clear responsibilities, requirements for testing and monitoring, documentation and – for material applications – independent review.
From an asset manager's perspective, this may give rise to practical questions such as:
• What role does the model actually play in the investment process?
• Which data drive its results?
• How are robustness and stability monitored?
• When is a model output challenged or overridden?
• How are model changes controlled?
• Is there a clear fallback if the model fails or produces implausible results?
• Who has professional responsibility for its use and monitoring?
These are not general regulatory requirements for every portfolio. They are practical questions arising from the significance of the relevant model within the investment process.
Conclusion
AI is not only changing how investment signals are generated. It is also changing which risks need to be monitored within the investment process.
For portfolio managers, it is therefore not enough to ask: How good is the model? Just as important is: How robust are the data, monitoring, explainability and governance surrounding the model?
This is precisely where Portfolio Management and Risk Management intersect.
Peak Compliance AG specialises in Compliance and Risk Management outsourcing solutions for portfolio managers and managers of collective assets in Switzerland and Liechtenstein.
Status: September 2026.
