In this introductory piece, I described the ability of artificial intelligence (AI) to process vast and complex data-sets with remarkable speed and precision. Such strengths can apply in different ways across different asset classes and so, over the coming months, we will move on to consider each of these separately – starting with equities.
For equity investors in particular, AI brings advantages that are worth exploring, especially when we distinguish between predictive AI – or machine learning and generative AI – versus large language models (LLMs) and agent-based systems.
Often, AI-based models are compared to factor-based systematic strategies – and given the relative infancy of AI in the investment narrative and our love of comparative assessments, you can see the logic here. Yet the real question is not whether AI replaces factors – it is whether it can help us model equity markets in a more flexible and forward-looking way.
Let’s start with a quick recap: linear versus non-linear relationships and how they might affect share prices. Linear factors operate in a ‘straight’ correlated line – if the level of a factor increases or decreases by 1%, we assume the affected stock price will also move by a proportionate amount.
Traditional linear models rely on smoothly evolving relationships between stock sensitivities and features such as momentum, size and value and assume these will remain stable in different market conditions.
Non-linear effects are less predictable. For one thing, a 1% change in a driver may have little impact in one environment and a much larger impact in another. Market behaviour also evolves across regimes – volatility, liquidity, policy expectations and investor sentiment can all influence how signals behave.
Predictive AI – adding value
This is where predictive AI can add value. Machine learning techniques are designed to handle high-dimensional data and can model non-linearities and interactions without requiring us to pre-specify the exact functional form.
When applied carefully, these methods can extract incremental predictive information from large sets of characteristics and macro variables, particularly by capturing interactions that are difficult to spot or model using subjective or traditional ‘quant’ methods.
To illustrate why moving beyond strict linearity can matter, consider threshold effects. A modest momentum rally lasting a few days may appear insignificant – but, once stocks hit a certain threshold in level or accumulated gains, the market takes note and may, for instance, exacerbate the upwards move. Flexible models can, in principle, adapt more easily to these kinds of conditional dynamics.
“Predictive AI helps us work with large, complex datasets and capture non-linear, evolving relationships. Generative AI helps us interpret unstructured information and move towards more systematic, proactive analysis.
Investors do need to use their data in a very judicious way – not just kickstart an AI model and hope for the best. The risk of overfitting – building models that describe the past extremely well but fail in live markets – is real.”
Another potential advantage is being able to react to situations in which the playing field changes very quickly. If sentiment turns sharply against a segment such as technology stocks, a linear model will seek validation in a long history of data and it might take a long time for its response to play out. Machine learning frameworks can be re-estimated more frequently and incorporate a broader range of contemporaneous signals.
That said, you do need to use your data in a very judicious way – not just kickstart an AI model and hope for the best. The risk of overfitting – building models that describe the past extremely well but fail in live markets – is real. Robust validation, disciplined feature selection and strong governance are essential. AI is a powerful tool, but it is not a shortcut. In short, you cannot simply data-mine.
So far, we have discussed predictive AI applied to structured data: prices, fundamentals, macro indicators and alternative datasets organised in tables. Equity markets are also influenced by unstructured information, however – earnings call transcripts, regulatory filings, company guidance and broader newsflow.
Generative AI – a complementary role
Here, generative AI, including LLMs, plays a complementary role. LLMs are not primarily designed as return-forecasting engines – their comparative strength lies in interpreting and structuring text at scale. They can extract sentiment, detect shifts in tone, identify emerging themes and convert qualitative disclosures into structured inputs. These can then be incorporated into broader predictive frameworks.
Generative AI also opens the door to more proactive analytical workflows. In an ‘agentic’ set-up, systems can continuously monitor news, filings and transcripts, flag unusual developments, summarise drivers and propose structured hypotheses for human review.
In practice today, this works best within a ‘human-in-the-loop’ framework, where portfolio managers retain oversight and decision authority. Fully autonomous trading driven by such systems would require very robust guardrails, risk controls and governance structures.
The implication is not that machines replace judgement – rather, predictive AI can help model complex, evolving numerical relationships, while generative AI can help digest the growing volume of qualitative information that influences markets. Together, they can assist portfolio and wealth managers in reducing the gap between information and actionable insight.
Connecting stock behaviour to portfolio resilience
The dynamics of share price trajectories extend beyond individual performance. They hinge on characteristics that link stocks with similar traits, such as size, profitability or balance-sheet strength. Traditional models provide useful approximations, but they inevitably involve trade-offs.
Effective portfolio construction – quant-based or otherwise – often highlights diversification as a core risk-management tool. Yet focusing purely on the quantity of risk can obscure the nature of the risks being taken. A forward-looking approach, supported by AI, enables portfolio and wealth managers to incorporate a wider range of data sources and update risk assessments more frequently than traditional portfolio rebalancing approaches.
Governance remains central. We often discuss AI in the context of signal generation, but robust model governance is just as important in portfolio construction and risk control. A future reliant solely on models for building and executing portfolios is neither realistic nor desirable. Professional human judgement is essential in selecting models, monitoring their performance and intervening during periods of market stress.
Portfolio construction, on the other hand, can be falsely reassuring – forecasting risk is generally easier than predicting returns. It would be easier to estimate the volatility range of AstraZeneca, for example, with greater confidence than to forecast its future share price return. This creates a misplaced sense that building portfolios is easy when, of course, it is not.
If a model estimates portfolio risk at 10% but realised volatility rises materially, a decision-maker must evaluate the situation and make a call. Is the model wrong or possibly wrong or is it right? And is the market behaving abnormally? We have clearly witnessed the latter scenario since the pandemic.
In equity investing, AI should therefore be viewed not as a black box, but as a toolkit. Predictive AI helps us work with large, complex datasets and capture non-linear, evolving relationships. Generative AI helps us interpret unstructured information and move towards more systematic, proactive analysis.
Used responsibly and governed carefully, together they can enhance equity selection, strengthen portfolio construction and improve resilience – while keeping human expertise firmly at the centre of the process.
Giovanni Beliossi is head of investment strategies at Axyon AI
