From Reference Portfolio to Asset Allocation
An Agentic AI Framework for Connecting the Mandate, Risk Target, and Portfolio Decisions
In my other articles, I have argued that institutional portfolio construction should follow a deliberate sequence.
It begins with the mandate.
The mandate defines the problem the institution is trying to solve. Risk appetite and risk capacity then help determine how much investment risk the institution is willing and able to sustain. The risk target turns that judgment into something measurable. The reference portfolio provides a simple and transparent expression of that chosen level of long-term market risk.
From there, a factor lens helps us look beyond asset-class labels to understand the economic risks and return sources actually embedded in the portfolio.
Ultimately, these concepts must lead to practical outcomes. The actual portfolio still has to be built.
This raises a key question I have been thinking about for some time:
Could we design a decision process that maintains a clear line of sight from the mandate and risk target, through the reference portfolio and views on risk premia, all the way to an implementable asset allocation?
Recently, I explored whether agentic AI could help structure this process.
Not by allowing AI to make investment decisions. Instead, specialised agents can organise different types of analysis, challenge assumptions, maintain consistency across the investment process, and provide decision-makers with a more disciplined path from institutional purpose to portfolio implementation.
The result is a relatively simple conceptual framework. This framework is not intended as a production-ready investment system. It explores how investment architecture and emerging technology can be integrated while keeping human judgment and accountability central.
Starting Before the Portfolio
A portfolio construction system should not start by asking an optimiser for asset weights. It should begin with a thorough understanding of the institution.
For a pension fund, that means considering the mandate, liabilities, funded position, benefit payments, contribution flexibility, liquidity requirements, investment horizon, governance arrangements, and tolerance for adverse outcomes.
Other institutions will have different considerations. An endowment, sovereign wealth fund, central bank, insurance company, or foundation may have very different objectives and constraints.
This distinction is important because the same investment opportunity can have different value and risk depending on the institution.
In an agentic framework, the first layer, therefore, organises institutional inputs and identifies key constraints for portfolio design.
This does not mean asking AI to determine the institution’s risk appetite. Risk appetite is fundamentally a governance judgment. It reflects what an institution is willing to accept in pursuit of its objectives.
AI can, however, help translate that discussion into analysis.
For example, a risk-target agent could evaluate alternative levels of investment risk and illustrate their consequences under different market and liability scenarios. It could examine potential drawdowns, liquidity requirements, contribution implications, funded-status outcomes, or other measures relevant to the institution.
It could ask:
- What happens if the institution takes more risk?
- What happens if it takes less?
- Which downside outcomes become difficult to sustain?
- How much liquidity might be required under stress?
- Does the proposed risk level remain consistent with the mandate and institutional capacity?
The agent can organise the evidence and highlight the trade-offs. It should not make the final decision. The approved risk target remains a human governance decision.
This distinction is important because the risk target is not just a model output; it is the quantitative expression of an institutional choice.
From Risk Target to Reference Portfolio
Once the institution has determined how much risk it is prepared to own, that risk needs an anchor.
That is the role of the reference portfolio.
As I discussed in From Risk Target to Reference Portfolio, I see the reference portfolio as distinct from a detailed strategic asset allocation. Its purpose is to provide a relatively simple, transparent expression of the institution’s long-term market risk.
The reference portfolio should therefore remain relatively stable. It should not change in response to shifts in expected returns, new investment products, or trends in asset classes. If the mandate, liability structure, or long-term risk target changes materially, the reference portfolio may need to be revisited. Short-term market fluctuations should not prompt changes to the reference portfolio.
Within an agentic system, this establishes an important boundary. Downstream agents should not optimise away the reference portfolio. It serves as the anchor for evaluating their recommendations.
The key question arises once that anchor is established:
Given the reference portfolio and approved risk bounds, how should the institution respond to new opportunities while maintaining focus on its chosen risks?
This is where the framework adopts a more agentic approach.
Building the Framework Across Investment and Technology
This project also became a collaboration across two disciplines.
I designed the investment architecture: what the system should start with, which investment judgments needed to remain distinct, how the reference portfolio should anchor decisions, how risk-premium views should be formed and challenged, how factor exposures should be translated back into investments, where risk should challenge the proposed portfolio, and where human decision rights needed to remain explicit.
I collaborated with my husband, who works in technology, to translate the investment logic into an agentic architecture. His perspective prompted me to reconsider orchestration, information flows, agent dependencies, validation, and system communication. It was an unexpectedly useful combination.
Investment and technology architectures address related challenges but approach them from different perspectives. An investment process may be conceptually sound but difficult to operationalise. Conversely, a technically elegant system may process information efficiently but fail to preserve investment intent.
Integrating these perspectives reinforced a principle that became central to the framework:
The objective is not to automate investment judgment, but to design a better system to support it.
The Macro Agent Frames the Environment
Once the policy foundation is established, the macro agent helps describe the environment in which portfolio decisions are being made.
Its role is not to forecast the future with false precision. Instead, it provides a structured approach to classifying the economic environment and identifying forces that may affect portfolio risks.
These might include:
- growth,
- inflation,
- real interest rates,
- credit conditions,
- market liquidity,
- fiscal conditions,
- currency dynamics, and
- geopolitical fragmentation.
The macro agent’s purpose is to frame the opportunity set, not to dictate the portfolio.
This distinction is important.
A macro view can become overly dominant if subsequent analyses simply inherit its conclusions. When everyone shares the same initial assumption, consensus may become just confirmation.
Therefore, the next layer of agents should operate with some independence.
Risk-Premium Agents Should Form Their Own Views
The next layer consists of specialised risk-premium and factor agents.
In the prototype, I used six broad areas, though the specific number is less important than the overall architecture. Institutions can define these dimensions based on their mandate and portfolio.
The agents might examine areas such as:
- growth risk,
- inflation risk,
- real-rate risk,
- credit risk,
- liquidity risk, and
- currency risk.
Importantly, a factor is not necessarily a risk premium.
An institution may care about inflation, for example, because inflation affects its liabilities even if it does not expect to earn a persistent premium simply for being exposed to inflation.
Agents therefore need to answer several questions.
- Is the exposure relevant because the institution needs it?
- Is there a compensated risk premium associated with it?
- Is that premium currently attractive?
- And is the portfolio already exposed to the same underlying risk through other investments?
Each agent can evaluate a defined set of questions:
- Is the expected premium attractive relative to history?
- What evidence supports the view?
- How uncertain is the estimate?
- What would invalidate the thesis?
- How does the exposure behave under stress?
- Is the risk already embedded elsewhere in the portfolio?
- Does increasing it improve the total portfolio or deepen an existing concentration?
- Is the exposure desirable because of expected return, because of the institution’s liabilities, or both?
These agents should not simply inherit the macro agent’s conclusions. They should evaluate their evidence independently. The views should be reconciled only after these independent assessments are complete.
That separation is important because sound portfolio decisions require both synthesis and challenge.
From Independent Views to CIO Synthesis
The CIO synthesis layer can then compare the independent assessments with the broader macro classification.
- Where do they agree?
- Where do they disagree?
- Is an apparently attractive credit premium consistent with deteriorating liquidity conditions?
- Does an inflation view make sense given what the real-rate agent is seeing?
- Are multiple recommendations ultimately expressions of the same underlying growth exposure?
Contradictions are not necessarily errors. They can contain useful information.
The purpose of synthesis is not to force consensus among agents, but to understand why views differ and what those differences mean for the total portfolio.
This is one area where I see considerable potential for agentic AI.
- It can make assumptions more explicit.
- It can separate evidence from conclusions.
- It can record where judgments differ.
- And it can make the path from an initial view to a portfolio recommendation easier to reconstruct later.
Transparency is essential because investment processes should be reviewable, not merely repeatable.
Looking Through Asset Classes
Once views have been formed, another problem emerges.
Investment views may be expressed in terms of economic factors and risk premia, but portfolios are implemented through actual assets. Those two languages are not the same.
Public equities, private equity, infrastructure, real estate, credit, nominal bonds, inflation-linked bonds, hedge funds, and cash are not pure exposures. Each contains combinations of underlying risks.
Private equity, for example, may include substantial growth risk, leverage, sector concentration, and liquidity exposure. Infrastructure may provide inflation sensitivity in some circumstances, but it can also contain equity, duration, regulatory, and leverage risk. Credit can contain both interest-rate and economic-growth exposure.
This is why diversification by asset class does not always result in diversification by risk.
The framework therefore requires an asset-decomposition layer. Its purpose is to map implementable assets back to the economic exposures that motivated the investment views in the first place.
This forms the bridge from factor judgment to asset allocation. Without this bridge, an institution may have a sophisticated macro framework but hold a portfolio with exposures inconsistent with its stated convictions.
Optimisation Is a Translation Mechanism
Once the expected premia, factor exposures, covariance assumptions, reference portfolio, and implementation constraints have been specified, optimisation can play a useful role.
However, the optimiser should have a deliberately limited mandate.
It is not being asked:
What is the optimal portfolio?
This question is too unconstrained.
Instead, it is being asked something closer to:
Given the institution’s approved risk anchor, current views, implementation constraints, and risk limits, what set of portfolio adjustments most effectively expresses those views?
That is a very different problem.
- The reference portfolio remains the starting point.
- The risk and factor agents provide views.
- The asset-decomposition layer translates those views into implementable exposures.
- The optimiser identifies potential deviations from the anchor within approved limits.
The result is therefore not a new policy portfolio generated from scratch. It is a constrained portfolio recommendation.
This distinction is important, particularly for long-horizon institutions. Optimisation should support the investment process rather than quietly redefine it.
As a practical test, I also implemented a simplified version of the framework using liquid ETFs as proxies for the underlying asset exposures and ran historical backtests.
The results were encouraging.
The exercise indicated that the framework could translate changes in risk-premium views into portfolio adjustments while remaining anchored to the reference portfolio and specified constraints. The resulting allocations were generally intuitive when examined through the lens of underlying economic exposures.
The backtest, however, should not be interpreted as evidence that the framework can reliably predict future returns. The ETF universe, historical period, transaction assumptions, model specifications, and parameter choices all influence results.
The purpose was more modest. The goal was to test whether the investment logic could withstand real market data and investable instruments.
On that basis, the results were promising.
Risk Should Challenge the Portfolio Before Approval
The optimiser’s answer should not move directly to decision-makers.
The next step should be a dedicated risk challenge. In the prototype, this is referred to as the CRO agent.
Its role is not simply to verify that formal limits have been met. It should ask what the proposed allocation actually does to the institution.
For example:
- How does expected drawdown change?
- What happens to expected shortfall?
- How do stress losses change?
- Does liquidity usage increase materially?
- Has the portfolio become more dependent on leverage?
- Does funded-status risk increase?
- Are contribution outcomes affected?
- Has currency exposure changed?
- Is diversification improving, or are several positions expressing the same underlying risk?
- Is the additional expected return genuinely associated with a risk the institution intends to own?
At this stage, risk management contributes directly to portfolio construction rather than serving solely as a post-decision control.
The purpose of risk is not to veto investment. It is to improve the quality of the decision.
The CIO Agent Organises the Decision
After the risk challenge, the CIO synthesis agent can bring the analysis together for human decision-makers.
Again, the AI agent should not substitute for the CIO. Its role is to organise the investment case.
A useful decision memo should make the logic clear:
- What changed?
- What views were formed?
- Where did the agents disagree?
- What allocation change is being proposed?
- Which risks are being increased or reduced?
- How does the recommendation relate to the reference portfolio?
- What are the expected benefits?
- What are the principal risks?
- What would cause the institution to reconsider the decision?
- Which assumptions matter most?
An effective investment memo should do more than present a conclusion. It should make the reasoning behind the recommendation clear.
Validation May Be the Most Important Agent
During prototype development, one part of the architecture became more important than expected.
Validation.
AI significantly accelerates analysis, but increased speed introduces new risks.
During the process, coding errors, logic errors, inconsistencies between assumptions, and outputs that appeared plausible but did not reflect the investment intent were encountered.
The system could produce an answer quickly. It, however, could not always determine whether the answer made investment sense. That assessment required human judgment.
For this reason, validation should not be the final step but a layer integrated throughout the entire process.
A validation or quality-control agent could test for:
- inconsistent assumptions,
- broken data dependencies,
- unintended constraint binding,
- double-counting of factor exposures,
- optimiser exploitation of weak assumptions,
- contradictions across agents,
- outputs inconsistent with the reference portfolio,
- recommendations that technically satisfy a rule while violating its economic intent.
This is particularly important in institutional investing. A model can obey the letter of a constraint and still produce a portfolio that makes little sense. Validation must therefore extend beyond code checks. It must also test investment logic.
Human Decision Rights Remain Explicit
One feature of the architecture I would not automate.
Accountability.
Different agents can analyse the mandate, organise institutional information, test candidate risk targets, classify economic regimes, evaluate risk premia, map factor exposures, optimise portfolios, conduct risk challenges, and prepare decision materials.
But they should not own the final decisions.
The board or relevant governing body retains responsibility for the mandate, risk appetite, and appropriate delegations.
- Human decision-makers approve the risk target.
- The investment organisation decides how to implement the institution’s investment strategy relative to the reference portfolio.
- The CIO and delegated investment teams remain accountable for portfolio decisions.
- Risk functions provide independent challenge.
Governance determines who has authority to decide. This matters particularly when investing fiduciary capital.
An AI system cannot be accountable to beneficiaries. It cannot fully understand institutional history, political context, stakeholder expectations, organisational capacity, or the consequences of a decision in the way accountable human leaders are required to do.
Technology can support the decision architecture. It cannot, however, assume ownership of the mandate.
A More Explicit Investment System
Taken together, the conceptual flow looks something like this:
![]() |
| Illustration |
Validation should be integrated throughout the entire process, not only at the end. The process should also periodically revisit policy.
The reference portfolio should not change because of market discomfort or new asset-class trends. However, over longer horizons, the institution should assess whether changes in mandate, liabilities, liquidity requirements, risk capacity, or governance arrangements warrant reconsidering its risk target and reference portfolio.
This distinction preserves both stability and adaptability.
What Agentic AI Actually Adds
The most significant promise of agentic AI in institutional investing is not improved prediction.
- Markets remain uncertain.
- Expected returns remain difficult to estimate.
- Models remain imperfect.
AI does not change those realities. Its practical value may lie elsewhere.
- It can make a complex investment process more explicit.
- It can reduce the manual effort involved in gathering and reconciling information.
- It can force assumptions into the open.
- It can maintain a record of why different judgments were made.
- It can identify inconsistencies earlier.
- It can create structured challenge between different analytical perspectives.
- It can help translate investment analysis into decision materials much more quickly.
- And, perhaps most importantly, it can help preserve the connection between decisions that are too often discussed separately.
The danger in asset allocation is not only choosing the wrong asset mix.
It is losing the line of sight from the mandate to the risk target, from the risk target to the reference portfolio, from the reference portfolio to the risks and premia the institution intends to own, and from those intentions to the exposures actually embedded in the implemented portfolio.
Agentic AI may help maintain that line of sight. However, this is only useful if the investment architecture is sound.
Technology cannot correct an institution that has not decided what it is trying to achieve, how much risk it can sustain, which risks it has a reason to own, or who is accountable for the decisions.
That is why, even in an increasingly agentic investment process, the philosophy does not change.
- Begin with the mandate.
- Translate purpose into an explicit risk target.
- Use the reference portfolio as the long-term anchor.
- Understand the risks beneath the asset classes.
- Challenge the portfolio before acting.
- And keep the final decision with accountable people.
The tools may change. The responsibility does not.
This is a conceptual framework developed as an exploration of how agentic AI could support institutional portfolio design. I designed the investment architecture and worked with my husband, who works in technology, on translating that investment logic into an agentic workflow. I also tested a simplified version using liquid ETFs and historical market data. The framework is intentionally simplified and is not intended to represent a production investment system or a claim about future investment performance.
* Images generated using Claude

Comments
Post a Comment