Generative AI can give you a competitive edge in the market. But you need to know how to use it. There is a right way to utilize the technology – and a wrong. That can be the difference between outperforming the market or making an ill-advised decision.
On November 30, 2022, a blogpost appeared on the website of the non-profit American AI-research lab OpenAI. About 4.000 words long, and the world hasn't felt the same since. The title of the blog post? “Introducing ChatGPT”.
Two months and 100 million new users later it’s fair to say something had changed forever.
2023 has been the generative AI’s time to shine. And the hype has been well-deserved. For the first time the ability to utilize this new computing-powered technology and with Large Language Models (LLM) has been brought to the masses with a wealth of new applications and use cases to follow.
Any self-respecting company has ever since been asking itself how to harness its qualities. And for an industry like institutional asset management, where asset managers are continually vying for a technological edge - of course asset managers were among the first to look at potential applications.
However, generative AI isn’t without its weaknesses.
LLMs can produce so-called hallucinations, and these aren’t great when you’re looking for a straight answer. Especially if you’re taking important business decisions based on results that are factually incorrect.
The thing is LLM-powered tools like ChatGPT aren’t better than the data put into the LLM behind the chatbot. Also, these have no built-in guarantee that they deliver factual correct answers, or even consistent answers with every query. This is due to the probabilistic nature of their design.
If you use AI to predict how you should put together your portfolio and rely on a chat assistant who has limited – possibly outdated – knowledge of the world and is biased towards certain things, then it will be an awful use case.
Research engineer director, SimCorp
That can result in bias in responses, nonsensical answers, and data-poisoning. So, knowing the limitations of technology is crucial if you’re looking for a competitive advantage.
In other words, there’s a right way and a wrong way to use generative AI.
“You must understand the technology is not infallible. In fact, it’s theoretically provable, they’re fallible, and make mistakes. You can make a model that predicts share prices, but you aren’t guaranteed that it’s correct. In many cases it will be wrong because share prices move all the time. When you look into some use cases from a user's perspective, the risk of making wrong prediction is simply too high if you want to use this for decision making,” says Allan Peter Engsig-Karup.
Learn more about SimCorps use of automation and machine learning here:
How SimCorp uses AI and machine learning
Sustainability data and compliance
Together with Clarity AI SimCorp offers an AI-powered solution for sustainability data and compliance. The offering is based on methodologies built from the ground up, that leverage advanced technology to streamline data collection and processing, reducing errors associated with manual work.
Intelligent Document Processing
SimCorp's machine learning-based offering Intelligent Document Processing automates the extraction of unstructured data from private market funds, which has historically been a manual and labor-intensive task for investors. The reason is the nature of the data from alternative investments, which investors typically receive in various forms and formats, typically PDFs.
AI-augmented, fixed income portfolio management
Through cloud-based connectivity from SimCorp, IntelliBonds delivers its Portfolio Assist AI-augmented platform. It adapts to clients’ specific investment strategies using machine learning, with virtual AI assistants that test and monitor investment strategies.
With a background as an associate professor in Scientific Computing at Denmark Technological University, DTU, he has spent years researching in Computational Mathematics and modern scientific computing paradigms and connecting these developments to engineering applications.
Allan Peter Engsig-Karup has considerable experience in transforming fundamental research into new innovative technology and for years, he has been working within the realm of Machine Learning and AI.
Besides his background within the realms of higher education Allan Peter Engsig-Karup works as a research engineer director at SimCorp.
In other words, he has a deep insight into the do’s and don’ts when it comes to reaping the benefits of generative AI.
As he points out, to use the technology in a commercial context, you need it to deliver faultless results and mitigate ‘failure modes’ of emerging technology on a consistent basis. This has been the guiding principle for SimCorp's approach to leveraging the power of generative AI. For the technology to be right most of the time isn’t always sufficient.
“If you use AI to predict how you should put together your portfolio and rely on a chat assistant who has limited – possibly outdated – knowledge of the world and is biased towards certain things, then it will be an awful use case. Sometimes it will go well and other times it will go really, really bad,” says Engsig-Karup.
In the last few years, it has become much easier to extract data automatically from documents than previously with combinations of tools. AI is a key component because it can find patterns in paragraphs or tables and extract data correctly.
Research engineer director, SimCorp
So, the obvious question is how to use the technology right? How can you unleash the power of generative AI in a way that improves efficiency, reduces costs, and gives you an edge over the competition?
A big advantage of the LLM technology is that they are excellent at common tasks such as machine translation, summarization, auto-completion of text, which is useful in programming, etc. Importantly, they can be used to understand the human language and match it with different sources of information offering vast opportunities for tapping into any knowledge base, e.g., processing and searching for just the right information.
Solutions like compliance rules for investment prospectus and automation of workflows play to the very strength of the technology. And that is the ability to process enormous amounts of data in very little time and identify automatically what is most like a compliance rule and match these within the SimCorp investment management platform for a seamless user experience.
It might not sound as sexy as the thought of an AI-powered chat assistant that is able to pick the winners on the market, but that doesn’t mean that generative AI can’t have a profound impact on your business.
“For me, the good use case overall is that it supports you doing what you want to do. If you have a problem that you want to solve, then you have to see AI as a tool that can help do it.”
“This means that you can use the technology to solve problems that you may not have been able to easily before. An example of this could be document processing at SimCorp. In the last few years, it has become much easier to extract data automatically from documents than previously with combinations of tools. AI is a key component because it can find patterns in paragraphs or tables and extract data correctly,” says Engsig-Karup.
Since 2016, he has been at the forefront when it comes to innovation and creating a technological edge at SimCorp.
He sees a lot of players on the market touting solutions based on chatbots like ChatGPT. However, when it comes to usage of machine learning and generative AI Engsig-Karup sees important shortcomings.
Machine learning and AI can be extremely effective at solving specific tasks. But that only goes for a certain type of problem, and you need to be aware what kind of questions it can answer meaningfully. At the end it all comes down to the data that are used to train the generative AI models as well as the data that you want to process with AI. There has to be some alignment between these sources for the technology to excel in performance.
“If you look at ChatGPT, it is trained on data up until September 2021 and knows a lot about everything in the world up to that point. But its output is not necessarily factually correct. There are many use cases where you cannot actually rely on the information even if it looks like it is correct. AI models are predictive machines and will always be able to predict something whether correct or not,” says Engsig-Krarup.
AI models are basically trained on historical data. Technically, they are really bad at extrapolating, i.e., predicting something into the future based on historical data unless the same pattern repeats itself.
Research engineer director, SimCorp
He continues: “If you don't know which data sources it was trained on or that it can be wrong, and still use it, it may result in misinformation. If you want to let it recommend a stock that has been moving up for a while, then the model will show that it can only go up, and when it falls, it predicts incorrectly.”
“AI models are basically trained on historical data. Technically, they are really bad at extrapolating, i.e., predicting something into the future based on historical data unless the same pattern repeats itself. And in how many use cases is it the same pattern that repeats itself forever? It’s probably limited.”
Therefore, unlocking the power of generative AI hinges on finding explicit business outcomes that plays to the strengths of the technology.
An example of this is SimCorps Intelligent Document Processing solutions – and the business effects are tangible.
According to Alkymi Analyst Survey 55 percent of analysts spend more than one hour a day searching for and copy-pasting information. Applying machine learning to the process means you’re able to automate a highly manual and time-consuming process.
The AI technology can make a decidedly big difference when it comes to creating structure in otherwise unstructured data. That is in other words the right way to use generative AI.
Any technique that enables computers to mimic human intelligence, using logic, if-then rules, decision trees, and machine learning (including deep learning).
A subset of AI that includes abstruse statistical techniques that enable machines to improve at tasks with experience. The category includes deep learning.
Models and tools designed to create new content, such as text, images, videos, music, or algorithms and code.