AI: Changing the discretionary asset management space
Introduction
Who has had this experience with their child? My baby daughter is 18 months old. When she was 12 months old I drew her a giraffe. Not a very good giraffe. A month ago I showed her a clip of a giraffe mother licking her newborn. She immediately said giraffe! That is from a single poor training example, whereas today’s AI require 100,000s of training examples. This is the dream of general intelligence AI. As we have not yet achieved this goal in AI nor are likely to soon I shall speak today rather AI agents helping with skilled but nonetheless boring and repetitive tasks.
I agree with Max Tegmark that there is no secret sauce for general intelligence or consciousness. We shall crack both these in time.
In the beginning
I am Justin Solms. This is my story of how AI has affected my life and my company. I’m the founder and Chief Investment Officer at Index Solutions, a low-cost asset manager in partnership with the Itransact investment platform.
I started off my profession in the 1990’s as a weapon guidance engineer subcontracted to the likes of Denel (formerly Armscor). Certainly the most fun project I ever participated in was a millimetre wave radar-guided, cruise-missile, seeker-head. Dr Graham Brooker, (now of the University Sydney) and I, believed we needed a state-of-the-art artificial intelligence to understand the radar terrain image and compare the extracted terrain features to an on-board map. We achieved a 5 meter by 2.5 meter target accuracy a 100% of the time; independently of the point of launch.
This was during a time we today jokingly refer to as the mythical AI winter of 1984 to 2006. Neural networks of that time were notoriously unreliable and difficult to train for certain tasks such as this one.
Instead I lovingly hand-coded a rules based, radar-image-processing AI for feature extraction, with statistical inference for map feature comparison. The resulting output was a map reference vector with a covariance matrix for the map reference accuracy (which was critically important to the inertial navigation system). We were severely constrained by computer processing power so we used specialized processors operating in parallel.
Today we have access to orders more processing power, and data, and most importantly, labelled data. Whilst the results would be excellent even today, the AI we use would be radically different, and would involve modern machine-learning and certainly deep-learning for the map feature extraction. There is still definite application for this more classical AI technology.
In 1999 I became invited to become a founding member of Peregrine Quant. At that stage there was major distrust of neural network based AI asset allocation. The literature was abundant enough about this. The task of Dr. Tim Gebbie (now of the University of Cape Town) and I was to develop tactical asset allocation using cutting edge quantitative techniques.
The statistical machine learning algorithms we chose at the time involved adaptive learning machines. We borrowed heavily from my experience as an engineer in online adaptive learning systems for the market forecasting component.
We bolted onto this the best asset allocation techniques that the research literature had to offer. This produced excellent results.
A point that was entirely missed was the cost of the team we ran and the resulting fee load on the client and the manager. The system required experts such as ourselves to maintain and run it.
Today in investment management we find ourselves in the midst of a fee war where costs are becoming more and more transparent and fees are being relentlessly driven down. At the same time investors are realizing that they are paying fees for products that underperform the market. As you know, and as our research supports, in South Africa, about 80% of managers underperform their benchmark, whilst charging fees. One of the 5 big banks has an bill-board advertisement in the Cape Town airport arrivals hall that claims that over 80% of managers under-perform the actual market. Unfortunately, the less than 20% managers who beat the market are not the same year-on-year.
The 20% is this an illusive and moving target. Chasing the latest outperforming manager shall incur significant losses in the form of transaction costs. For this reason it is simply better to invest in the market. The question is: “what is the market?”. We use machine learning to determine this.
We now face pressure to deliver performance in line with markets, and at a low total investment cost to the client. This makes sense for everyone as statistically you are more likely to be a client than an asset manager. One arrives at the realisation that it will be the low cost manager that successfully scales their service to the larger number of clients; that will have the higher likelihood of being profitable in the future.
Low fees are possible through scaling
Performance is possible through maximum extraction of information from the market and then acting upon this information with almost zero error.
Because of this we find ourselves ever increasingly turning toward automation of tasks that were not previously not possible for machines to perform and which were solely the domain of expert humans.
The promise of AI is to change all this.
Index Solutions together with the Itransact retail investment platform operates in the retail world. This implies lots of clients with lots of different needs.
So when we consider automation in our business we have four main drives toward AI. And these are:
- Asset Allocation (Already fully automated – Machine learning)
- Robotic advice (Mostly automated – Rules based)
- Automated reporting (Partly automated – rules based)
- Lifestyle services (New research & development)
A word about AI terms
- Rules based – Rules are hand coded into an algorithm and its goal is specified or specific.
- Machine learning – NVIDIA defines machine learning as the practice of using algorithms to parse data, learn from it, and then make a determination or prediction about something in the world.
- Neural networks – These are shallow information processing networks of neurons which bear a superficial resemblance to human cortical neurons. These were the earlier attempts at machine cognition and are prone to failure.
- Deep learning neural networks – Deep layers of neural networks. Each layer provides greater abstraction of features than the previous layer, superficially resembling the human neocortex.
The theory goes that as we evolved from simpler beings into humans we added layers to the frontal cortex of our brains, enabling us to achieve deeper levels of abstraction of our ideas. The key breakthrough in deep learning is the large number of layers. This too enables greater abstraction of the input features to arrive at an output.
Unfortunately deep learning is a black box, albeit a very successful black box. Fathoming the manner of how an individual network does what it does is an area of intense research.
Asset Allocation
At Index Solutions we use machine learning algorithms for security selection and risk based asset allocation. Our process results in portfolios of broad market tracking indices that are risk adjusted to client needs. This satisfies our goal of delivering the market; as explained earlier.
Or process also ensures we rid our products of human cognitive biases. An example of a cognitive bias is cherry picking of evidence in support one’s preconceived belief system. This is one of the most dangerous cognitive biases in my opinion. It is also death to portfolio performance. We do not second guess the algorithm, rather we focus on maximum extraction of information from the data with careful selection of the trade-off between algorithm bias and algorithm variance. We test and test and test again, often with tens of millions of test runs.
When we market our services we have to be careful to whom we speak as this technology is not trusted by everyone despite our excellent track record. We are often asked about our views on the market. We humans don’t have any views, we only have data, algorithms and outcome probabilities; which is hard to swallow for traditional investors and investment managers.
Robotic advice
Our Itransact investment platform has built and deployed a type II robo-advisor as an experimental sales tool.
I will not cover chatbots here as these we believe they are an add-on to robo-advisors and a different technology covered later on in the conference.
Investopedia defines robo-advisors as digital platforms that provide automated, algorithm-driven financial planning services with little to no human supervision.
Deloitte estimates between 2.2 and 3.7 trillion USD in assets will be managed with the support of Robo-Advisory services in 2020. By the year 2025 this figure is expected to rise to over 16.0 USD trillion.
Robo-advisors encourage the self-management of one’s financials. Robo-advisors provide information in a totally different way that doesn’t require a deep financial skill from the investor.
A study by Deloitte of German robo-advisors finds there are four kinds of robo-advisor, each building on the other:
- An online questionnaire leads to a product selection.
- Investments are of funds with real managers and risk based allocation. This is where our company is currently at.
- Allocation adjustments made by algorithm with a rule set.
- Fully automated asset allocation with self-learning algorithms. Still in the initial phase.
The caveat here is Turnově and timing. Timing markets is notoriously difficult and has significant risk of drawdown. In our experience turnover leads to portfolio cost increases, losses, and taxation events. It also isn’t clear how regulators are going to treat these robo-advisors once the complaints start to roll in.
We believe it’s still early for these machines and we have pegged our robo-advisor at level two. We will wait and see how well robo-advisors perform in the real world. We can afford to wait as in South African uptake for robo-advisors is still weak.
There exist issues of trust in these machines and the press will certainly be following their performance as their track record evolves.
There exists the issue of client confidentiality and the POPI act. Robo-advisor providers may soon discover that they require retail-banking level security and privacy measures.
Robo-advisors are the future, either becoming agents alongside real advisors, or standalone. Their potential to further reduce investment fees, we believe, will win the day for the stand-alone robo-advisor.
The threat of job loss to the independent fund advisor is uncertain. The real barrier is not technology, but rather trust and regulation. And perhaps a generational change too as the expected adopters, the younger investors, are still building wealth. Most of my peers in my generation are not ready to accept robotic advice.
Automated reporting
AI journalists are writing their own reports, AI authors are writing realistic horror fiction. AI technology is rapidly approaching maturity for automatic financial report writing.
Index Solutions and the Itransact retail platform’s research focuses on personalised report writing for each individual client.
Report writing is arguably one of the most boring tasks for any investment manager. In large organisations the accuracy of the report must be independently verified, corrected and signed off by a senior person. This is an expensive use of resources, even with the automated office tools. Writing individual reports for 1000’s of clients would be prohibitive.
An experienced report writer can complete a small to medium size report in half an hour. Longer reports with imaginative content take much more time to prepare. Current AI report writers can do the job in 10 minutes or less. AI is currently unable to distinguish between dull and interesting and may sometimes make dubious conclusions. One AI report writer claimed that Brexit would be a tailwind for the UK economy.
These problems will soon be entirely solved. But generating novel content and theme, as is the practice in retail report writing, may be some years away and will continue rely on humans.
There are obvious marketing benefits to personalised reports.
We have reached the stage of producing a personalized fact sheet as a cover page to the statement for each individual client, however this is a far from true AI report writing and is an active area of research for us. We are keeping close eye on off the shelf products.
You are fortunate if report writing is simply an extra task you wish to pass off to someone or something else. In the near future this will mean the obsolescence of jobs for many low-paid people doing report writing.
An interesting and current development is artificial AI. Pinterest is using crowd sourced humans evaluate and write. These are “humans pretending to be robots pretending to be humans”.
Lifestyle services
Amazon is well known for its product recommender services. Whist these can be sometimes quite annoying and incorrect, relevance certainly has increased in recent years.
I am still advised to buy 50 shades of grey and Powerful and Feminine. Amazon in all ways demonstrates a deep belief that I am female for which I am truly astonished and annoyed and seemingly powerless to rectify. I have bought over 20 books covering global wars and terror conflicts in Iraq and Afghanistan, and they still don’t get it.
Financial service providers who are smart enough to structure their robo-advisors and client platforms to collect as much relevant data as possible will be in the unique position to understand their client’s needs in ways never humanly possible at scale. The internet of things will also become essential.
Fraud and spam detection are trusted and successful AI (ML) technologies. These technologies rely on detection of outlying features instead of routine.
Applying similar algorithms to collected client data, AI would be capable of detecting client distress or outlying needs, such as being in financial trouble, requiring legal help, or having a health issue.
Partnering asset management and retail services with preferred service providers in the medical, health & well-being, fitness, financial and legal industries, would provide a business model for AI driven recommendation of personalised lifestyle services to current clients.
Finder fees would accrue to the recommender from these preferred service providers. This would be similar to Google add revenue generators at the top of search result pages.
The consequence of this cannot be understated.
As the costs and fees of financial services are further driven down, a business model whereby financial services are provided at low cost or even free; would become viable. This would be financed from the revenue from recommendation of lifestyle services.
This could be the final frontier in zero cost financial services. This would move the focus to a future of integrated low cost services, transforming asset management as we know it.