Forecasting Engine for Financial Portfolio Planning

Regulatory Framework

The regulations governing portfolio decisions converge on a duty to act in the client's best interest, to document why a recommendation was made, and to demonstrate that risk was bounded by something other than the recommender's discretion. SEC Regulation Best Interest requires broker-dealers to act in the retail customer's best interest at the time of a recommendation, with care, disclosure, conflict, and compliance obligations whose discharge must be evidenced. FINRA Rule 2111 requires that a recommendation be suitable based on the customer's investment profile and that reasonable diligence include consideration of potential risks and rewards. The Investment Advisers Act of 1940 imposes a fiduciary duty on registered advisers that the SEC has interpreted as encompassing duties of care and loyalty, with the duty of care including a duty to provide advice in the client's best interest, to seek best execution, and to provide ongoing monitoring.

In Europe, MiFID II Article 24 sets out general principles requiring firms to act honestly, fairly, and professionally, while Article 25 requires assessments of suitability and appropriateness with a written record of how the recommendation matches the client's knowledge, experience, financial situation, and objectives. The FCA's Conduct of Business Sourcebook implements and extends these requirements in the United Kingdom, with COBS 9A on suitability for MiFID business and the Consumer Duty layering further obligations to deliver good outcomes. Annex III of the EU AI Act classifies certain AI systems used in creditworthiness, life and health insurance pricing, and related financial activities as high-risk; portfolio tools that influence access to financial services or that automate decisions affecting natural persons increasingly fall within scope and trigger the Article 9 risk management, Article 10 data governance, Article 12 logging, and Article 14 human oversight obligations. Basel III/IV, while primarily a prudential regime for banks, shapes the constraints on bank-affiliated portfolio activity through capital, leverage, and liquidity rules whose internal model governance under SR 11-7 in the United States and equivalent supervisory expectations elsewhere requires structural validation of every input that drives a regulated number.

The common thread across these regimes is that portfolio decisions must be defensible by reference to a documented process that bounds risk, considers alternatives, validates suitability, and produces evidence available for supervisory review.

Architectural Requirement

An AI system that participates in portfolio decisions under these regimes must exhibit four architectural properties. It must maintain alternatives as first-class objects, so that the consideration of options required by suitability and best-interest analysis is structurally visible rather than implied by the existence of a chosen recommendation. It must contain speculative analysis so that exploration of aggressive or unconventional positions does not influence the live portfolio until validation has been completed. It must validate every promotion to action against a typed set of constraints derived from the client profile, the firm's investment policy, the regulatory limits applicable to the account type, and the prudential rules where relevant. And it must produce lineage that ties every executed allocation change to the analysis, scenarios, validations, and authorizations that preceded it.

These properties define a planning surface rather than a recommendation surface. A recommendation is a point output; a plan is a structured representation of conditional intent that supervisors and clients can examine. The architectural requirement is that the AI system expose the plan, not only the chosen action, with the plan's conditions, alternatives, validations, and promotions all accessible as typed data.

An architecture that does not exhibit these properties produces evidence of suitability and best interest only by retrospective narrative, in which the firm explains, after a decision has been executed, what alternatives were considered and why the chosen action was preferred. Retrospective narrative is precisely the form of evidence that current rule-making is moving away from, in favor of contemporaneous structural evidence that does not depend on the firm's own account of its reasoning.

Why Procedural Compliance Fails

Procedural compliance for portfolio AI consists of policy documents that describe how the AI is supposed to be used, supervisory routines that sample its outputs, and reconciliation reports that compare its recommendations to outcomes. Each of these is a procedural commitment to behavior that the technical layer does not enforce structurally, and each fails under the rules cited above in distinct ways.

It fails on the alternatives requirement. Suitability analysis under FINRA Rule 2111 and MiFID II Article 25 contemplates that the firm has considered alternatives and selected the one that best matches the client's profile. A model that produces a single recommendation without a structured representation of the alternatives it implicitly rejected cannot evidence that consideration. The supervisory record will show the chosen action and the model inputs, but the alternatives exist only in the model's internal computation and are not available for review.

It fails on the containment requirement. Without structural containment, every analysis the model performs has the potential to influence live recommendations, because there is no typed boundary between exploration and decision. Under EU AI Act Article 9 risk management and Article 14 human oversight, this absence of containment translates into an inability to demonstrate that the human-in-the-loop has access to a stable artifact on which to exercise oversight; the artifact moves as the model recomputes.

It fails on the validation requirement. Procedural validation consists of policies that say which constraints apply and reviews that check whether they were honored. Reg BI's care obligation, COBS Consumer Duty's good outcomes test, and Basel internal model governance all require that constraints be enforced at the moment of action, not only checked after the fact. A planning surface that promotes recommendations to action without typed validation cannot meet that requirement structurally.

It fails on the lineage requirement. SEC and FINRA examinations, FCA supervisory visits, and EU AI Act conformity assessments increasingly ask for the chain of analysis that justifies a specific position. Procedural compliance answers this with logs and emails; structural compliance answers it with a planning graph whose nodes carry the analysis, conditions, validations, and authorizations as fields. Logs and emails are reconstructable but not authoritative; planning graphs are authoritative because they are the operational object on which the decision was made.

What AQ Primitive Provides

The AQ forecasting engine implements the planning surface as a directed graph of governed branches. A base case branch reflects the current allocation under expected conditions. Alternative branches represent the portfolio's response to specific scenarios such as rising interest rates, equity drawdown, geopolitical disruption, sector rotation, or liquidity stress. Each branch contains a complete rebalancing plan: positions to adjust, magnitudes, sequence, and expected cost, together with the conditions under which the plan becomes eligible.

Each branch is wrapped in a containment boundary that prevents speculative analysis from affecting the live portfolio. Within the boundary, the agent can simulate aggressive tactical positions, evaluate the consequences of stress scenarios, and explore alternatives that would be unsuitable to execute. The containment is structural rather than procedural; the engine does not permit a contained branch to emit execution instructions.

Promotion from a contained branch to executable status is gated by validation against typed constraints. The constraints encode the client's risk tolerance and objectives drawn from the suitability profile, the firm's investment policy and concentration limits, the regulatory limits applicable to the account type, and the prudential constraints where the activity sits within a regulated balance sheet. The validation is a structural step whose output is a typed authorization, not a discretionary judgment that the analyst then defends in narrative.

Conditional promotion enables graduated response as evidence accumulates. As market indicators align with a scenario branch, the engine incrementally promotes elements of that branch's plan, executing early steps while keeping the full rebalancing contained until the trend is confirmed. Branch dormancy preserves the planning structures of scenarios whose probability has declined, so that re-emergence of similar conditions activates a known plan rather than triggering replanning under pressure. Lineage is produced continuously: every analysis, every validation, every promotion, and every execution is recorded as a typed entry on the relevant branch, with cryptographic chaining that ties entries to their predecessors and to the agents that produced them.

Compliance Mapping

SEC Regulation Best Interest's care obligation maps to the validation gate that constrains promotion: the gate enforces that a recommendation be the result of a process that considered alternatives and matched the customer's profile, with the planning graph providing the structural evidence. The disclosure obligation maps to the branch metadata, which can be rendered into client-facing disclosures derived from the same artifact that drove the decision. The compliance obligation maps to the engine as a whole, which is the operational surface that implements the policies and procedures Reg BI requires.

FINRA Rule 2111 suitability maps to the branch structure: the alternatives that the rule contemplates exist as branches, and the chosen branch's promotion through validation evidences that the recommendation was suitable for the customer's profile. The Investment Advisers Act fiduciary duty of care maps to the same structure, with ongoing monitoring satisfied by the engine's continuous evaluation of branch eligibility as market conditions evolve. MiFID II Article 24 maps to the engine's overall conduct, while Article 25 suitability and appropriateness map to the validation gate and the lineage of the suitability assessment recorded on the branches. FCA COBS 9A and Consumer Duty map to the same artifacts, with the good outcomes test supported by the engine's record of which alternatives were considered and why the executed plan was selected.

EU AI Act Annex III high-risk obligations map to specific structural artifacts: Article 9 risk management to the branch validation framework, Article 10 data governance to the inputs recorded in branch lineage, Article 12 logging to the lineage entries themselves, and Article 14 human oversight to the contained-branch artifact on which the human reviewer can act without the model's continuing computation moving the target. Basel III/IV internal model governance, where applicable, maps to the validation gate as a control on inputs to regulated capital and risk numbers, with SR 11-7 model risk management satisfied by the engine's structural validation rather than by procedural review.

Adoption Pathway

Adoption proceeds in three phases, calibrated so that the firm's existing portfolio process continues to operate while the planning surface is introduced. In the first phase, the firm models its current investment policies, suitability frameworks, and regulatory constraints as typed validation rules and configures the forecasting engine to consume them. The engine runs in shadow alongside existing tools, producing planning graphs whose validation outcomes are reviewed against the firm's manual processes. The phase ends when the engine's validation produces results consistent with the firm's documented policy, with discrepancies resolved either by correcting the engine's configuration or by surfacing latent ambiguity in the policy itself.

In the second phase, selected portfolios or strategies are migrated to the engine as the system of record for planning. The validation gate becomes authoritative: rebalancing actions are executed only when the gate authorizes promotion, and the planning graph becomes the artifact on which suitability, best interest, and conduct evidence is generated. The migration is per-strategy and reversible; a strategy can revert to prior tooling if a regression is observed. Throughout the phase, the firm accumulates structural lineage that becomes available immediately for examinations, audits, and supervisory visits, even before all strategies have migrated.

In the third phase, the engine becomes the firmwide planning surface for portfolio decisions within the chosen scope. Regulatory change is then addressed by updating typed validation rules rather than by retraining staff or rewriting middleware. When new regimes emerge, whether further EU AI Act guidance, additional Annex III scope, or revised Basel rules, the firm responds by configuring the engine, and the response is auditable as a configuration change rather than as a process redesign. The engine's planning graphs become the durable record of how the firm formed and revised its investment intent under successive regulatory regimes, and the firm's compliance posture rests on structural evidence rather than on the persuasive force of retrospective narrative.