From Reference Portfolio to Asset Allocation

An agentic framework for translating risk-premium views into actionable portfolio tilts

Image: generated by Claude


A reference portfolio serves as the starting point for portfolio design, not its conclusion.
For a long-horizon asset owner, the reference portfolio should express the institution’s core risk appetite. It defines the broad level of risk the fund is willing to take in order to meet its mandate. It is not meant to answer every allocation question, and it should not change every time the market environment changes.
However, the actual investment portfolio still has to be built.
This is where the more complex work begins.
The key question is not simply, “What should the SAA be?” Instead, given a reference portfolio and defined risk bounds, how should the institution translate macro regime and risk-premium views into an actionable asset allocation?
This is the challenge I have been considering.
Recently, I sketched a simple agentic AI framework for doing exactly that: starting with the Investment Policy Statement and reference portfolio, assessing the macro regime, forming risk-premium views, decomposing asset classes into factor exposures, and using optimisation to derive active tilts within approved risk bounds.
The framework includes distinct agents: a macro agent, risk-premium agents, a CRO risk report, and a CIO synthesis agent. This division of labour helps separate types of judgment that are often combined prematurely.
Importantly, the agents provide recommendations but do not make final decisions.
Decision-making remains the responsibility of people.
This distinction is critical. In institutional investing, particularly with fiduciary capital, accountability cannot be delegated to a model. AI can organise evidence, test assumptions, identify inconsistencies, draft analysis, and compare alternatives. However, it cannot own the mandate, fully understand the institutional context, or be accountable to beneficiaries as a board, CIO, or investment committee can.
The underlying philosophy remains unchanged, even as the workflow becomes more agentic.
The process must still begin with the mandate.
In this framework, the Investment Policy Statement is foundational. It defines the reference portfolio, risk limits, governance constraints, and review cycle. The reference portfolio serves as the anchor, and the optimisation process operates strictly within the mandate.
The macro agent comes next.
Its role is not to forecast the future with undue precision, but to classify the environment in a disciplined manner. Factors such as inflation, growth, real rates, credit conditions, liquidity, fiscal pressure, and geopolitical fragmentation influence which risk premia are likely to be rewarded or vulnerable. The macro regime does not provide the answer, but it helps frame the opportunity set.
The risk-premium agents operate at the next level.
In the framework, I use six factor risk-premium agents, though the exact number is less important than the principle. The goal is to assess compensated sources of return separately, such as growth risk, inflation risk, real-rate risk, credit risk, liquidity risk, currency risk, or any factor structure relevant to the institution.
Each agent should address a defined set of questions.
  • Is this risk premium attractive relative to history?
  • Is it being rewarded in the current regime?
  • How uncertain is the estimate?
  • What would make the view wrong?
  • How does this factor behave under stress?
  • Does adding it improve the total portfolio, or does it create hidden concentration?
This is where agentic AI adds value.
Not because it can predict which risk premium will outperform, as it cannot.
Its value lies in making investment judgments more explicit, consistent, and reviewable. It separates assumptions from conclusions, documents the rationale, and enforces a challenge process before optimisation.
It can also significantly reduce the time required.
This is a significant advantage. Traditionally, building such a framework requires weeks of manual effort, including gathering inputs, checking assumptions, reconciling definitions, mapping exposures, coding, reviewing outputs, and drafting decision materials. AI can accelerate much of this work.
However, increased speed does not equate to full automation.
While building this framework, I identified and corrected several errors, including coding and logic issues, as well as outputs that appeared plausible but did not align with investment intent. AI accelerated the process, but it could not independently determine which errors were significant. Human judgment, validation, and correction remained essential.
This experience reinforced an important lesson.
Agentic AI is powerful but requires validation and quality control. A knowledgeable human is needed to identify when outputs are incorrect, incomplete, or misleading.
A validation or quality-control agent may be among the most critical components of the framework.
Its role extends beyond verifying code functionality. It should assess internal consistency of assumptions, unintended constraint bindings, optimiser exploitation of input weaknesses, double-counting of factor exposures, and whether recommendations violate the policy's intent, even if they meet formal constraints.
This is where the workflow becomes more interesting.
  • The macro agent may frame the regime.
  • The risk-premium agents may form views.
  • The asset decomposition layer may translate factor exposures into asset-class inputs.
  • The optimiser may solve for active tilts.
  • The CRO agent may challenge risk, stress, drawdown, and liquidity.
  • The CIO agent may synthesise the case for decision-makers.
The validation agent should oversee the entire process, ensuring all components are coherent. This is a significant responsibility. In many respects, this role maintains the integrity of the framework.
The subsequent step is asset decomposition.
This is critical because asset classes are not pure building blocks. Public equity, private equity, infrastructure, real estate, credit, nominal bonds, inflation-linked bonds, hedge funds, and cash all contain mixtures of underlying factor exposures. If the investment view is expressed in terms of risk premia, the portfolio must be mapped back to asset classes before it can be implemented.
In other words, the process moves from factor judgment to asset allocation.
That bridge matters.
Without it, a fund can have a sophisticated macro view and still end up with an asset allocation that does not reflect the intended exposures. Or it can add private markets, credit, inflation-linked assets, or hedge funds without fully understanding how much growth risk, liquidity risk, credit beta, duration, or currency exposure it has actually taken on at the total-fund level.
Once expected returns, factor exposures, covariance, and constraints are specified, optimisation can do its job.
In the sketch, I describe this as a CVXPY active-tilt solve. The optimiser is not producing a new policy portfolio from scratch. It is solving for active deviations from the reference portfolio, subject to risk bounds and implementation constraints.
That distinction matters.
  • The reference portfolio sets the long-term risk anchor.
  • The risk-premium agents form active views.
  • The asset decomposition maps those views into implementable asset classes.
  • The optimiser derives the allocation tilts.
This approach differs from traditional mean-variance optimisation applied indiscriminately to asset classes. It is a mandate-constrained, factor-aware process for translating risk-premium views into portfolio positions.
But even then, the output should not go straight to approval.
A CRO risk report should challenge the proposed allocation. How does the tilt change VaR, expected shortfall, drawdown, stress losses, liquidity usage, funded-status risk, contribution risk, and currency exposure? Does the portfolio remain within the approved risk bounds? Is the improvement in expected return coming from compensated risk, or from exposures the institution does not actually want?
At this stage, risk should inform investment judgment, rather than serve solely as a post-hoc control.
Finally, the CIO agent synthesises the analysis into a board memo.
Again, the point is not to replace the CIO. The point is to organise the logic. What changed? What is the proposed tilt? Which risk premia are being increased or reduced? How does the allocation remain anchored to the reference portfolio? What are the key risks? What would cause the institution to revisit the decision?
An effective board memo should not only report outcomes but also clarify the decision-making process.
The dashed line back to policy review is also important. The reference portfolio and risk bounds should not be constantly adjusted in response to market noise. But they do need periodic review. Every few years, the institution should ask whether the reference portfolio still reflects the mandate, liability structure, liquidity needs, and governance capacity.
That is the distinction I would want to preserve.
  • The reference portfolio is the strategic anchor.
  • The active tilt process is the adaptive layer.
  • The optimiser is the translation mechanism.
  • The validation process is the quality-control layer.
  • The governance process is what keeps the whole system accountable.
This is where agentic AI can provide significant value for institutional investors. Not as an allocation oracle. Not as a black box. Not as a substitute for investment judgment.
Its value lies in establishing a more disciplined workflow: beginning with the mandate, explicitly evaluating risk premia, mapping factor views to asset exposures, stress-testing allocations, validating logic, and documenting trade-offs for decision-makers.
The danger in asset allocation is not only choosing the wrong asset mix.
It is a loss of transparency among the mandate, reference portfolio, assumed risk premia, and the implemented portfolio.
Agentic AI can help preserve that line of sight.
  • It can save time.
  • It can make assumptions more visible.
  • It can expose errors earlier.
  • It can force more consistency across the workflow.
  • It can help transform analysis into actionable decision materials.
However, it cannot eliminate the need for human judgment, nor should it.
Regardless of whether agentic AI is used, the philosophy remains: begin with the mandate, understand the risks, challenge recommendations, and ensure decisions are made through accountable human governance.
The tools may change.
The responsibility does not.

Comments

Post a Comment

Popular posts from this blog

Stockholm, Market Timing, and Broadening the Total Portfolio Conversation

Earlier this week, I visited Stockholm to participate in the Nordic Pension Tour. Thank you to Mirjam Guldemond, Sophie Baker, and Pensions & Investments for organizing an excellent program. It was one of those trips that reminds you why these conversations matter. Stockholm itself helped: the weather was beautiful, the city was full of that calm Scandinavian clarity, and even a short visit to the Nobel Prize Museum made it feel like a place where ideas still carry real weight. Before the World Pension Summit, I visited AP3, Alecta, and FTN. Each institution provided a unique perspective on how long-term asset owners think about their investment. At Alecta, I noted the central role of benchmarks. For liability-aware institutions, benchmarks are deeply embedded in the investment process. While this can be helpful, it raises an important question: when does a benchmark clarify the mandate, and when does it begin to replace it? At FTN, together with AP7, we discussed defined contribut...
Read more

Leadership Starts with Motivation

The older I get, the more I think leadership is less about authority and more about motivation. A leader’s job is not just to set direction, make decisions, or evaluate performance. Those things matter, of course. But if the people around you no longer feel motivated to contribute, improve, and stay, then a lot of the rest begins to break down. Motivation has never meant just compensation. Pay matters. Performance-based rewards matter. Fairness matters. But motivation also comes from something more basic: being treated with respect, being seen as an individual, and having a working relationship with your manager that feels human rather than mechanical. Every workplace is, in the end, a human system. People may stay for a while because of prestige, compensation, or obligation. But very few people do their best work for long in an environment where they feel invisible, unfairly treated, or repeatedly drained by the people around them. That is why I think good leadership requires more tha...
Read more

The Mandate Comes First

Image
One of the more enduring rewards of this year’s ALTSLA, once again, was the people. The quality of the conferences is usually assessed by the quality of their panels, the relevance of the themes, or the prominence of the speakers. All of that matters. But what stays with me most is usually the less formal part — the conversations between sessions, the chats over coffee, the passing comments that turn into more substantive discussions. For all the attention this industry gives to ideas, it is people who ultimately shape them. That was on my mind as I rode the train from Los Angeles to San Diego after the conference, enjoying the improbably warm weather compared with the East Coast. Over time, ALTSLA has become something of an annual ritual, and the two days of rest that follow it have become part of that rhythm as well — an opportunity to step away from the noise, reflect, and reset. Looking out at the ocean, I found myself returning to a question that strikes me as basic, but not basic...
Read more