Automated asset allocation stands on the brink of a seismic shift in which AI will transform the competitive landscape by scaling efficiency and exponentially enhancing performance. However, the prevalent unease surrounding the opacity of black-box AI models is hindering its broader acceptance. This use-case explores the potential and viability of AI solutions for automated asset allocation, focusing on systems that are not only efficacious but also transparent, trustworthy, and compliant.

Asset allocation, a cornerstone of investment management, guides investors in distributing their capital across diverse asset classes. A popular adage advises the following:
Let every man divide his money into three parts, and invest a third in land, a third in business, and a third let him keep by him in reserve.”
In today’s interconnected financial landscape, investment funds seek diversified returns by broadening their exposure to various asset classes, geographies, and styles. The automated asset allocation sector has witnessed substantial growth, driven by investment firms’ need to minimise operational costs and automate processes. Automation through rule-based algorithms allowed a significantly smaller number of human investors to manage an expansive array of mutual funds and other pooled investment vehicles, enabling a significantly higher daily volume of investment. This automation is made possible by the conversion of client-investment manager agreements into algorithmic configurations that are combined with other inputs such as subscriptions, redemptions, market indicators, and other economic variables, to automate trading activities. The investment manager’s role is twofold: (i) accurately translating the client agreement into the requisite algorithmic configuration and (ii) diligently overseeing the investment decisions to validate orders (a mandate for legal compliance) and rectify any algorithmic discrepancies or failures.
The landscape investment managers navigate is intricate, characterised by the management of numerous funds with thousands of daily cashflows, inclusive of both subscriptions (inflows) and redemptions (outflows). These cashflows, alongside the current investment position, market fluctuations across asset classes, jurisdiction-specific calendar variations, and the myriad stipulations embedded within investment management contracts, are used collectively to determine trading activities in the form of raised orders.
The investment manager must then review these orders, either approving them or intervening in cases of exceptions. The challenge intensifies with the sheer volume of orders necessitating review, the complexity of recalling multi-faceted investment management contracts, especially in large quantities, and the formidable task of correlating each order with its respective contractual clause keeping in account constantly changing data and market conditions. Moreover, the detection of errors stemming from order omissions further complicates the scenario, as inactivity can also be indicative of errors.

In the automated asset allocation, two potential applications of AI are immediately evident.

1. Contract to Configuration Translation: Post-client acquisition, contracts, articulated in legal terms, are translated into system configurations by investment managers and their teams. This translation, traditionally static barring contractual alterations or error discovery, stands to benefit significantly from AI integration. AI could be applied either to autonomously generate configurations from contract text or to critically evaluate manually crafted configurations, suggesting refinements and highlighting potential discrepancies based on its training and the contract’s content. While human oversight remains indispensable for this critical phase, AI promises substantial enhancements in speed and efficiency for this laborious process.

2. Configuration and Inputs to Order Creation: In this phase, the system autonomously proposes trading orders by combining the configuration with cash flows, prevailing market conditions—such as price fluctuations and restrictions like sanctions and holidays—and the current investment position. Presently, this phase is mechanised through manually constructed rules-based systems. However, the ever-increasing complexity of these systems inherently limits their scalability. In this context, AI solutions offer a compelling alternative. With AI systems’ capacity to autonomously discern patterns during training, they could effectively overcome these scalability constraints. However, this comes at the expense of losing the expert knowledge captured in existing rule-based systems.

The reluctance towards adoption of AI-based solutions emanates from various concerns, predominantly the opacity of black-box AI models. The complex, non-linear nature of high-performing modern AI algorithms, such as neural networks, renders a comprehensive and precise understanding of their inner workings and their individual predictions unattainable – they are black-boxes. This opacity impedes human oversight, making it virtually impossible to pinpoint discrepancies, assess risk, or validate model and prediction compliance.

For the automated asset allocation applications highlighted, the lack of AI explainability negates potential advantages. For instance, in the contract-to-configuration scenario, human reviewers would still need to meticulously peruse the entire contract to authenticate the AI-predicted configuration, effectively nullifying the algorithm’s efficiency gains. Similarly, in order prediction, the lack of AI transparency restricts human reviewers to merely corroborating AI predictions against those from rule-based models, once again, counteracting the potential benefits of AI. Coupled with the initial investment required for AI development and deployment, these concerns often lead stakeholders to opt for lower-risk alternatives.

Explainable AI (XAI), or AI systems with transparent models that ‘explain’ their inner workings and predictions in a human-interpretable manner, could provide the essential oversight for the outlined applications. Addressing the growing scepticism towards black-box models, the domain of XAI is witnessing notable advancements. Two predominant approaches to XAI currently exist:

  1. Explainability Techniques: These methods (such as LIME, SHAP) require post-processing of each prediction and only provide an estimate of the contribution of individual inputs. The lack of absolute certainty, combined with the substantial time required for post-processing, renders them unsuitable for our high-volume, high-velocity critical applications.
  2. Inherently Explainable AI Models: These models, often termed white-box models, offer both model transparency and human-interpretable prediction explanations. While historically, explainable models lagged their black-box counterparts in terms of predictive performance, recent developments in Neuro-symbolic AI promise to overcome this problem. Neuro-symbolic AI aims to seamlessly combine the predictive performance of neural networks with the structured reasoning of rules-based symbolic logic. At the forefront of this Neuro-symbolic AI paradigm is UMNAI’s Hybrid Intelligence. Hybrid Intelligence models not only match or surpass the performance benchmarks set by leading black-box models but also ensure complete transparency, predictability, and full auditability. These models concurrently deliver predictions together with detailed, precise explanations in real time, communicating the precise contribution and reasoning behind each input’s influence on the output prediction.
    Hybrid Intelligence profoundly simplifies AI’s application to asset allocation. Investment managers, equipped with this technology, can scrutinise each prediction in real-time, using the explanations to ascertain the suitability of an order or to propose alternative actions, keeping the human in the loop. The precise real-time explanations also pave the way for further automation and efficiency enhancements, as codifiable checks traditionally performed by investment or compliance managers on a prediction can be integrated into automated workflows that directly act upon the prediction and its explanation. Furthermore, the explanations, detailing the precise construction of a prediction from its inputs, also double as audit logs, which can be rendered tamperproof and archived systematically.
    Another benefit of Hybrid Intelligence is its capacity to train models on existing legacy rule-based systems concurrently with training data, preserving valuable expert knowledge learned over time while easing the transition to AI-based solutions.
    In these applications, a phased approach to AI deployment is always recommended, initially running an AI-based solution in parallel to existing rule-based systems. This strategy allows for comparative performance analysis and continuous AI solution refinement, fostering confidence prior to a full transition to AI.

The financial sector, characterised by its complexity and broad reach, is at a pivotal moment where it faces dual pressures: the demand for lower costs alongside increased accuracy, accountability, transparency from regulators and clients. This transparency requirement, coupled with competitive pressures to reduce management fees, stands in stark contrast to the rising costs associated with more sophisticated investment strategies and higher volumes of transactions. In this context, the integration of Hybrid Intelligence into automated asset allocation emerges as a compelling solution. It combines the processing power of AI with the nuanced judgment and adaptability of human oversight, aiming to streamline processes, reduce errors, and ensure decisions are transparent, understandable, and justifiable. By embracing this technological evolution, the financial sector is poised to enter a new era marked by increased efficiency, reduced risk, and enhanced trust. As Hybrid Intelligence and other AI technologies continue to advance, they promise to propel the investment management industry to a brighter future where the balance between lowering fees and managing increasing operational transparency is successfully achieved.