AI-driven portfolio construction opens the door to bespoke baskets

Artificial intelligence could give structured product issuers another way to build more tailored equity baskets, with portfolio construction and rebalancing handled by machine-learning models rather than fixed selection rules.

The core use case is building and optimising bespoke equity baskets - Frank Siu, Arabesque AI

AutoCIO, developed by Arabesque AI, uses machine-learning forecasts to help construct equity portfolios subject to specified constraints. The platform can build a basket from scratch or analyse an existing portfolio and recommend changes to its holdings and allocations.

Frank Siu, chief investment officer at Arabesque AI, said the platform’s primary purpose was portfolio construction rather than the development of any particular investment product.

“The core use case is building and optimising bespoke equity baskets,” he said. "How those portfolios are ultimately used, whether in structured products, exchange-traded funds (ETFs), derivatives, separately managed accounts or funds, is a decision for the user".

That distinction could be relevant to structured products, where the underlying basket can determine the exposure investors receive. AutoCIO offers a way to construct and maintain a portfolio before deciding how to package it, rather than starting with a product wrapper and working backwards.

However, the platform is not currently being used to manage structured products. Arabesque’s asset management business invests in single-stock equities and does not use derivatives or structured products.

Building and rebalancing baskets

AutoCIO can work with an existing portfolio as well as create a new one. Users can upload their holdings and ask the system to recommend changes to position sizes or allocations, including adding or removing stocks.

Siu (right) described the portfolios as dynamic because their composition can change over time. Rebalancing may involve adjusting existing weights or replacing constituents as forecasts and portfolio requirements change.

Most of the live client mandates he was aware of used equity baskets rebalanced every one to three months, although the platform can accommodate other frequencies.

The approach also allows users to impose constraints on the portfolio. These can include liquidity requirements, sector exposure and limits on individual positions, enabling the resulting basket to reflect specific investment objectives or operational requirements.

In a demonstration, AutoCIO constructed a global portfolio of 50 stocks focused on artificial intelligence and technology, subject to liquidity, sector and concentration constraints.

Apple received a forecast score of 62%. Siu described this as the model’s estimated likelihood that the stock would rise over the following few months. It is a model output rather than a guaranteed probability of a positive return.

The platform draws on a range of inputs, including financial and accounting measures, macroeconomic indicators, valuation metrics, technical indicators, analyst estimates and information from other asset classes.

Its forecasting models use machine-learning architectures rather than large language models.

“LLMs are used in the AI assistant and explainability functions, helping users interact with the system and understand its outputs,” he said

Applying the approach to existing baskets

The potential application to structured products can be considered through existing stock-basket strategies, where issuers or product providers may need to select constituents and determine their weights.

Two examples supplied for analysis were a Van Lanschot Kempen basket and a 30-stock KBC basket. In the Van Lanschot Kempen example, a simulation over approximately one and a half months in August 2026 indicated that the allocations recommended by AutoCIO would have performed about two percent better than an equal-weight version of the basket.

The result is limited to that historical period and comparison. It does not establish that the approach would outperform in future or that the same outcome would be achieved after transaction costs and other implementation considerations.

The KBC example included target weights and a maximum position size of eight percent. That limit was retained in the analysis to reflect the basket’s stated maximum, although the reason for the original constraint was not confirmed.

These examples illustrate how the platform can be applied to existing portfolios and their allocation rules. They do not demonstrate that AutoCIO is already being used to select underlyings for live structured products.

The portfolio comes before the wrapper

Siu said AutoCIO does not determine whether a portfolio should be used in a note, an actively managed certificate (AMC), an ETF or another investment vehicle. That decision depends on factors such as tax treatment, regulation and distribution considerations.

The distinction is important for any structured product application. A portfolio that can be constructed and rebalanced systematically still needs to be assessed for its suitability as an underlying, including whether its constituents and trading characteristics fit the intended product.

For issuers, the potential attraction is the ability to define a basket around particular objectives and constraints, then update it periodically instead of relying on a static allocation. But the product design, wrapper and terms remain separate decisions.

“The core use case is building and optimising bespoke equity baskets. How those portfolios are ultimately used, whether in structured products, ETFs, derivatives or other investment vehicles, is a decision for the user. AutoCIO focuses on the portfolio construction itself, rather than the wrapper.”

_{Image: Adobe Stock}

---