The emergence of large language models, such as GTPx from OpenAI, has captivated many with their extraordinary text and content generation capabilities. These probabilistic driven models have enabled organizations like Financial Simplicity to achieve more output with fewer resources by harnessing the power of generative artificial intelligence, realising that machines can actually do a lot of what humans thought that only humans could do.
In this paper, parallels are drawn between the utilization, benefits, even necessity of the latest ‘personalized GPTs’ innovations and similar trends in Financial Simplicity’s portfolio rebalancing algorithms business. The increased recognition that the role context plays in the respective domains is explored in the generation of ‘fit for purpose’ improved outcomes and experiences.
Probabilistic vs. Deterministic Models
In the world of artificial intelligence, two distinct paradigms stand out: probabilistic and deterministic models. Probabilistic models, like GPTs, generate text and content responses by predicting the most likely sequence of characters or outputs based on the given input. In contrast, deterministic models produce calculated responses using predefined datasets and rule-based algorithms.
One key difference lies in how these models approach data generation and decision-making. Probabilistic models introduce an element of probability and adaptability, while deterministic models rely on defined rules.
‘Scope Context’: Customization and Configuration
The notion of tailoring GPT’s or creating ‘personalised GPTs’ for specific subject domains is growing, with contextual tailoring becoming necessary to achieve the productivity levels required to deliver increased utility and maintain competitive positioning in specific domains.
In the same way, the notion of configuring portfolio rebalancing algorithms with data (often in the form of model portfolios) produces outputs specifically aligned to the views of the model portfolio author.
However, model portfolios and algorithms customised and configured for context are clearly more useful than those that are not, and yet in most cases today the audience is still generally treated as a cohort, meaning the algorithm design inherently assumes the nature of the user / recipient. The real life situations is that people have different preferences, and investment situations, preferences and pathways prior to rebalancing a portfolio that best are considered.
The Next Level: Personalisation and User Context
Including specific user/client context into algorithms is the next part of the puzzle. For over 20 years algorithms on sites such as Amazon have been able to generate mass personalised recommendations improving the individual consumer experience and related spend.
To improve the experience from generative AI tools and investment services, the algorithms and models behind them also benefit by personalising context and prompt input, which is increasingly the expectation.
OpenAI’s ChatGPT allows users to enter custom knowledge and user-specific details to improve contextual relevance for the specific user. Similarly, contemporary portfolio rebalancing algorithms should (if they wish to be most valuable) be configurable for an individual investor’s context.
Individual investor personalisation (e.g. timing & threshold instructions, long term portfolio directional, and individual investment specific directional constraints) can significantly improve contextual relevance, resulting in improved quality and relevant output, to the interactions between the algorithm, portfolio managers/advisers and their investor clients.
Fit For Purpose
In the same way as ChatGPT is seeking to bring best relevance of subjects with personalised GPTs, the goal of personalised rebalancing algorithms is enhancing service authenticity and improving investor experiences. In all these scenarios, the goal is to provide users with the most valuable and relevant information using technological intelligence.
The difference between being treated as part of a mass audience and receiving specialized attention in wealth management for an investor is the difference between a generic experience (which may be perceived of little value) and a personalised one, creating value and trust.
In the same way that a far too generic response from Chat GPT is not as valuable as something more tailored to each user, portfolio rebalancing algorithms that do not include the investor context sufficiently feel somewhat bland, and perhaps miss key points which could have detrimental or sub-optimal impact to investors.
Those portfolio managers/advisers directing client investments to ‘one size fits all’ rebalancing services risk limiting the client experience, possible sub-optimal outcomes or in the extreme case harm.
Today technological models are employed to elevate user experiences across a wide range of industries. This approach forms the core of many business innovations, including personalized GPTs and highly configurable investment algorithms.
On further examination, two distinct levels of adaptation and configuration are crucial to achieve superior outcomes:
a) Narrowing the ‘Scope’:
Customizing algorithms to focus on specific knowledge areas or use cases, increasing service provision relevance.
In the case of Chat GPT this is custom GPTs that prioritise knowledge in specific subject domains. The more knowledge loaded, the better the relevance of the output.
For portfolio rebalancing algorithms this specific subject domain knowledge is about setting model portfolios, scheduling, and timing instructions in the first instance, but over time can deliver better outcomes if configurable around investment security directional limits also (e.g. don’t buy a specific (perhaps overpriced)investment for new clients, whilst leaving existing clients with the investment holding it in their portfolio).
b) Considering the Requester/Investor:
Tailoring responses and recommendations based on individual user characteristics, context, and preferences is the next level of detail.
In the case of Chat GPT this is the ability to set ‘Custom Instructions’ to improve context and output.
For portfolio rebalancing algorithms this is about setting individual investor specific rules, preferences, and constraints to enhance rebalancing output increasing relevance for that investor.
This may be represented as below:
For custom GPTs
For portfolio rebalancing
The Importance of Alignment
The quality of the output in this conceptual model hinges on the alignment of all three components. For generative AI and portfolio rebalancing algorithms to deliver the highest service levels and most relevant output, these elements must work harmoniously together, and so they must be designed that way in the first place.
If rebalancing algorithms and models are not designed to accommodate the service provider (portfolio manager) and end requester (investor / adviser) context then this can lead to poorer outcomes resulting in reworking the output, introducing errors and taking time. An example of this is the need for seperate ‘pre-trade compliance’ processes which limits the ability to scale businesses.
A lack of alignment in rebalancing algorithms can lead to the following types of questions from investors and/or their advisers:
Q) Why is the rebalancing algorithm determining portfolio adjustments inconsistent with my client’s long-term objectives?
Q) Why is the rebalancing algorithm not recognising that the investor has a significant capital gain tax liability already and has realised additional unwanted capital gains?
Q) Why is the rebalancing algorithm indicating that an investor purchases an already overvalued investment when contributing cash ?
Q) Why is the rebalancing algorithm generating many insignificant adjustments for the portfolio value?
Q) How do I deal with the fact that I may have some holdings with considerable capital gains that I don’t wish to be realised?
Q) Who designed the rebalancing process, and what are the design principles and values inherit in the process?
Q) How do I instruct the rebalancing process to exclude investments that I never wish to hold in the portfolio?
Q) Are the design principles around this algorithm consistent with best outcome for my clients?
Disclosure and Due Diligence
In the same way that there are (in many cases mandatory) disclosures around attributes of investment products in the market, where there is a rebalancing component, there will likely be focus on increasing disclosures around the features and /or the types of rebalancing algorithms that investors and their advisers use so they can compare, contrast and assess suitability before the placement of client monies, not by finding out after the event.
In conclusion, the parallel between user contextual GPTs and personalised portfolio rebalancing algorithms lies in their pursuit of customisation and contextual relevance. Both approaches aim to elevate user experiences, providing tailored solutions that meet specific needs and foster stronger relationships.
As technology continues to evolve, aligning these elements becomes increasingly crucial for delivering sufficient and quality service and achieving the quest for better answers. This is becoming the norm in every industry and failure to do so can result in suboptimal outcomes or disappointment.
In a highly regulated industry like Financial Services, the onus and responsibility of those who advise clients into products and services that use algorithms to rebalance and adjust client portfolios is likely to come under greater scrutiny – rebalancing functions performed in regulated products may not necessarily mean that they provide the best investor outcome.
For those with regulatory responsibility for designing algorithms or advising clients into products and services where algorithms are a key component to delivering investor experiences, the need to understand and compare different providers will become increasingly important in due diligence. This may become as important, if not more important (based on possible outcome deviations), as understanding attributes such as fee levels or investment mandate restrictions.