From Investment Intent to Portfolio: The Rise of Conversational Portfolio Construction

Portfolio construction is entering a new phase. Over the past several decades, much of the innovation in portfolio construction has focused on the mathematics: richer objectives, more realistic measures of risk, increasingly sophisticated constraints, transaction costs, taxes, multi-asset class portfolios, and multi-period investment decisions. Those advances have made the underlying machinery much more powerful.

But the way investment professionals interact with that machinery has changed much less. Running a sophisticated optimization has traditionally required users to understand how to define the universe, select the appropriate benchmark and risk model, choose an objective, configure constraints, prepare the required data, and translate an investment idea into the terminology expected by the optimizer.

AI introduces another important dimension: a fundamentally different way to access quantitative portfolio construction.

This article is the first in a series exploring what we see as the next generation of the portfolio construction experience—one in which investors describe what they want to achieve in natural investment language, while model context protocol (MCP) enables the AI system to discover and use data, analytics, and mathematical optimization to formulate, solve, analyze, and refine the portfolio.

For example, here is an investor request in natural language: “Create a portfolio that replicates the S&P 500 using fewer than 50 securities while keeping sector exposures within 1% of the benchmark.”

For the investor, it is a single sentence. For a portfolio construction system, it represents a complete quantitative workflow.

A Simple Question, a Complex Quantitative Workflow

The investor’s request has several elements of an optimization problem.

• The S&P 500 defines the benchmark.

• Replication implies an objective such as minimizing active risk or tracking error.

• The portfolio may hold no more than 50 securities, introducing a cardinality constraint.

• Sector weights must remain close to the benchmark, introducing group-level bounds.

But important details are missing.

• What investment universe should be used?

• Which risk model should measure active risk?

• Are there minimum or maximum security weights?

• Is turnover constrained?

• Are transaction costs relevant?

• Should existing holdings be considered? How should cash be treated?

A traditional workflow requires the user to resolve those questions by navigating an application or writing code. A conversational portfolio construction workflow can instead help interpret the intent, determine the data needed, formulate the corresponding optimization, execute it, and return the resulting portfolio and analytics.

In practice, FPE Intelligence handles the gaps in one of two ways. FPE Intelligence is a coding agent for the FactSet Programmatic Environment (FPE), using dedicated context, tools, and skills to create full FPE workflows and artifacts from natural language prompts. 

• In a guided mode, the agent asks a few clarifying questions before running anything. For example: whether to start from cash or from an existing portfolio, which analysis date to use, which classification should define the sector bounds, and which risk model should measure active risk.

• In a streamlined mode, most choices are preselected with sensible defaults, the optimization runs immediately, and the defaults are reported alongside the result so they can be reviewed and changed.

Even with the natural-language request, the quantitative complexity remains intact. To make that point concrete, we submitted the investor request in FPE Intelligence, starting from cash. The optimizer returned a 48-security portfolio, respecting the 49-name limit implied by “fewer than 50,” with a predicted tracking error of 1.97% against the S&P 500. All sector weights stay within ±1% of the benchmark, with several of those bounds binding (Figure 1).

For the analysis date in this example, with fewer than a tenth of the index's names, the portfolio overweights the largest stocks (Apple at 8.66% versus 7.03% in the benchmark, and Microsoft at 7.54% versus 5.69%) and uses a small number of holdings as concentrated proxies for their sectors.

The accompanying risk reports show what that concentration implies for the same analysis date. Although the request said nothing about factors, the portfolio carries an active tilt toward larger companies (Size, +0.37) and away from value measures such as earnings yield and book-to-price (Figure 2). Risk is concentrated as well: NVIDIA alone accounts for about 22% of total portfolio risk, and the 10 largest contributors, dominated by mega-cap technology and semiconductor names, account for roughly 70% (Figure 3).

In our example, the user can ask the logical next question, such as “Neutralize the size exposure” or “Keep every position within 1% of its benchmark weight,” and the system will rerun the optimization with that requirement. Asking FPE Intelligence to neutralize the size exposure, for example, adds a constraint that keeps the active Size exposure within ±0.05.

In the re-optimized portfolio the Size tilt falls from +0.37 to +0.05, binding at the upper edge of the band; the portfolio uses all 49 permitted names, and tracking error rises from 1.97% to 2.22%. That quarter-point increase is the price of removing the large-cap tilt that a concentrated replication naturally picks up, and it is now an explicit, quantified trade-off the user can accept or reject.

The follow-up also reveals something the user did not ask about. In the concentrated portfolio, the size, liquidity, value, and profitability tilts shared a common source: overweighting a handful of mega-caps. Neutralizing size removed most of the others as well, leaving every style exposure within ±0.07 of the benchmark (Figure 4).

3 Components Have to Work Together

Conversational portfolio construction is therefore more than putting a chat interface in front of an optimizer. In our view, three components need to work together.

Component 1: The optimization engine as quantitative decision layer

The first component is the portfolio optimization engine. The investor may describe the desired outcome in natural language, but constructing the portfolio remains a mathematical optimization problem. The engine must simultaneously balance the objective against all applicable constraints and costs and return a portfolio that is mathematically consistent with the specification.

FactSet Portfolio Optimizer (FPO) is designed to construct multi-asset portfolios according to custom investment objectives while adhering to specified constraints and costs. Our white paper, Multi-Asset Class Portfolio Construction, describes support for equity, fixed-income, derivative, fund-level, and cross-asset workflows, including both mean-variance and scenario-based approaches to risk.

Component 2: Integrated data and analytics as the investment context layer

The second component is integrated investment data and analytics. Optimization, rarely a clean mathematical problem, begins with data: the investment universe, current holdings, a benchmark, security characteristics, prices, classifications, expected returns or signals, factor exposures, risk models, sustainability metrics, transaction-cost assumptions, and other portfolio analytics.

Gathering, mapping, cleaning, and aligning those inputs can be a substantial part of the workflow. A conversational experience is more useful when the system resolves and accesses those inputs directly rather than asking the user to prepare each one manually.

The FactSet Programmatic Environment provides a JupyterLab-based environment with programmatic access to FactSet data and analytical applications. The FPE overview emphasizes the ability to combine multi-asset class data, research, simulation, and portfolio construction within a cohesive quantitative workflow.

Component 3: The LLM as interaction and orchestration layer

The third component is the LLM-powered agent. Rather than numerically approximating the optimal portfolio, its role is to understand the user's investment intent, determine which information and tools are required, translate the request into a formal optimization specification, invoke the appropriate data and analytics, run the optimization, and help the user interpret and refine the result.

Within FPE Intelligence, that enables a conversational portfolio construction experience aligned with FactSet’s multi-layered AI stack, FactSet Intelligence, which combines data, agents, and a breadth of financial workflows.

Lowering the Barrier to Sophisticated Portfolio Construction

Perhaps the most important implication of this architecture is that it can make sophisticated portfolio construction accessible to a much broader group of investment professionals.

Portfolio optimization has traditionally required specialized knowledge before the first optimization can even be run. Users have needed to understand the structure of the problem, the available objective functions and risk measures, the required datasets, and the syntax or interface that configures the strategy. For quantitative portfolio managers, those concepts may be familiar. For many other investment professionals, the setup complexity can be a significant barrier to using optimization when their investment questions are often perfectly natural:

• “Can I reduce the number of holdings in this benchmark while maintaining similar risk characteristics?”

• “Can I increase exposure to quality without materially changing sector allocations?”

• “Can I reduce portfolio risk while keeping the portfolio close to my current holdings?”

• “Can I construct a bond portfolio with similar duration and credit characteristics but fewer securities?”

• “What happens if I allow 75 securities?”

• “Can I reduce tracking error further?”

• “Which constraints are binding?”

• “Now add a positive quality tilt.”

• “Keep the strategy but reduce turnover.”

These are portfolio construction problems, even when the person asking them does not think of them as optimization problems.

A conversational interface changes the starting point. Instead of requiring users to first learn how the optimization machinery works, it enables them to begin with the investment outcome they are trying to achieve. The system can then translate that intent into the appropriate data, analytics, objective, and constraints.

That does not eliminate the need for transparency or quantitative rigor. Professional users are still able to inspect the assumptions, benchmark, risk model, objective, constraints, and resulting portfolio. But they no longer need to master every technical detail before they explore what optimization can do for them.

The Optimizer Still Does the Hard Part

The simplicity of the interface should not be confused with simplicity of the underlying problem. For example, here is a richer request from an investor: “Construct a portfolio tilted toward momentum and profitability. Keep sector weights within 2% of the benchmark, active risk below 3%, turnover below 20%, and hold between 40 and 70 securities.”

A language model can understand that instruction and map the words to portfolio concepts. But understanding the request is not the same as solving it.

The requirements interact: a tighter sector constraint may limit the achievable factor tilt, and cardinality limits can make the problem combinatorial. The role of the optimizer is to consider the relationships simultaneously and determine the best feasible portfolio.

That is why generative AI can make optimization more important rather than less important, a point we will discuss in our next article.

From One Optimization to an Interactive Research Process

Conversational interaction also changes what happens after the first portfolio is constructed.

Traditional optimization workflows are often sequential: configure the strategy, run it, analyze the result, return to the settings, change an input, and run it again. A conversational workflow can make this process much more iterative.

After constructing the 48-stock S&P 500 replication portfolio, the user might simply continue:

• “What happens if I allow 75 securities?”

• “Can I reduce tracking error further?”

• “Which constraints are binding?”

• “Now add a positive quality tilt.”

• “Keep the strategy but reduce turnover.”

Each question or instruction modifies or investigates the portfolio construction problem that has already been established. The interaction begins to resemble a conversation with a quantitative portfolio construction specialist rather than interaction with a traditional optimization configuration screen.

The same idea extends naturally to diagnostics. If a requested portfolio cannot be constructed because the constraints are inconsistent, the useful answer is not simply that the optimization is infeasible. The useful questions are why it is infeasible, which requirements conflict, how far the problem is from feasibility, and what could be changed. We will return to this topic later in our series.

Beyond a Single Asset Class or Strategy

The conversational interface is not tied to one portfolio type. The same architecture can support very different investment problems because the optimization engine and data layer provide the underlying flexibility.

For example:

• An equity investor may want to replicate an index with fewer names, construct a minimum-volatility portfolio, target specific factor exposures, or reduce selected sustainability metrics.

• A fixed-income investor may want to replicate a bond benchmark while controlling duration, credit quality, sector exposures, issuer concentration, and transaction costs.

• A multi-asset investor may want to allocate across equities, fixed income, funds, and other instruments while controlling risk and allocation ranges across the entire portfolio.

Risk objectives can vary as well. Some problems are naturally expressed using volatility or tracking error. Others may benefit from scenario-based measures of downside risk such as Expected Tail Loss. The conversational layer gives the investor a common language for expressing the goal; the underlying data and optimization capabilities determine how that goal is represented mathematically.

The Next Generation of the Portfolio Construction Experience

For decades, advances in portfolio construction have largely focused on making the models more powerful. That work continues. More realistic objectives, richer constraints, better risk models, transaction costs, taxes, scenario analysis, and multi-period decisions all improve our ability to represent investment problems faithfully.

But the next generation of portfolio construction is also about making that power easier to access. Rather than thinking like an optimization engine, the portfolio manager should be able to describe the investment problem in the natural language in which the mandate is expressed. The system should help translate that intent into data, analytics, mathematics, and an executable workflow.

That can make experienced quantitative users faster. More importantly, it can broaden access to institutional-grade portfolio construction for professionals who historically may not have used optimization because the setup was too complex.

The sophistication remains, just in a simpler interface.

Where Our Article Series Goes Next

This article focused on the overall architecture and experience: natural language as a new interface to portfolio construction, supported by integrated investment data and a rigorous optimization engine.

The next article will examine why the rise of AI makes mathematical optimization more important, not less—particularly as investors use natural language to express increasingly rich combinations of objectives and constraints.

A subsequent article will focus on one of the most practical challenges in portfolio construction: infeasibility. We will explore how dedicated diagnostics and routines inside the optimization engine can identify the causes of an infeasible problem, while an LLM can help translate those diagnostics into understandable explanations and actionable suggestions for resolving the conflict.

A final article will turn from the conversation itself to what comes afterward. Once a workflow has settled, natural language is no longer the fastest way to run it. We will look at how the same conversational system can be asked to produce a purpose-built interface for a workflow—a panel with the inputs that matter and the reports that follow—while the specification underneath remains open to inspection and change.

To Learn More

FactSet Portfolio Optimizer: overview of FactSet's multi-asset class portfolio optimization capabilities.

FactSet Programmatic Environment: programmatic access to FactSet data and analytical applications in a JupyterLab-based quantitative research environment.

How to Select a Portfolio Optimizer that Supports Your Investment Priorities: earlier FactSet Insight discussion of optimizer capabilities, data integration, risk models, and platform flexibility.

FactSet Intelligence: multi-layered AI stack powering faster and smarter decisions for financial professionals.​

This blog post is for informational purposes only. The information contained in this blog post is not legal, tax, or investment advice. FactSet does not endorse or recommend any investments and assumes no liability for any consequence relating directly or indirectly to any action or inaction taken based on the information contained in this article.