The finance industry has had a lot of hype around the potential of machine learning as a tool to generate new and deeper investment insights.
The Technology session at Frontier’s annual conference aimed to demystify the topic by drawing on industry experts with machine learning experience. The panel were all in agreement that there is immense potential for innovation, however it is in early days in delivering the promise.
Michael Kollo, who leads a Quantitative Solutions team at a large industry super fund, said the reasons the finance industry will adopt machine learning will be to “forecast the world better, to understand it, to understand the truth better”.
Machine learning seeks to discover deeper insights from a vast amount of data, not replicating familiar human intelligence patterns. It has the potential to generate new insights for an industry that traditionally relies on a great deal of human intuition.
However, how the investment management industry embraces machine learning hasn’t always been so obvious. The panel shared their experience in potential applications ranging from highly hyped return generating models, to robots that challenge existing business models or investment strategies. The potential of scalability gains from generating insights from vast amounts of research was also raised.
“The reality I see at the moment, primarily outside of Australia, is that adoption is driven by scale and lower cost integration rather than discovering insights”, said Kollo.
Pitfalls and unique challenges were explored in detail to understand why the adoption is lagging its promise. Three key themes were discussed; existing issues with traditional quantitative/statistical methodologies, the data set and risk of building black box models.
The finance industry uses traditional quantitative/statistical techniques to explore historical data to predict future events. However, the reality is that these traditional forecasting approaches that the industry is based on, only “achieves 55% success rates and not the anticipated 80%. The reality is that it is still difficult to explain”, Kollo told the audience.
Modern techniques in technology aim to discover patterns on a larger scale, building on these foundations. Therein lies the heart of the problem which machine learning is not going to help with.
Frontier’s Maye Zhao said “these are traditional and classical problems in statistics. They are not new problems and they appear not because we use machine learning. These problems might never go away, and this is the reality we live in.”
The other reality of the domain is the unique data set that is more complex, dynamic and uncertain than other fields. Zhao reflected on her time building models to generate trades however had “to spend 40% of my time on data collection and data massage”. Artificial intelligence and machine learning is data intensive and hungry.
The final major challenge discussed is commonly referred to as black boxes. It is quite simply that the model becomes so complex the person building the model struggles to explain its logic.
As the industry is at the beginning of understanding the potential and value application of these new technologies the panel warned against jumping straight in as the answer, without consideration of these complexities. All speakers agreed that “these solutions require a great deal of human judgement and decision making to generate valuable insights for investment decision making”.