Daniele Grassi, CEO at Axyon AI
For many active managers across the world, the last few years have been challenging. Failure to meet their key benchmarks has become more regular, and investors have started to pursue passive funds instead.
In an attempt to recapture their edge and regain the faith of investors, active managers have increasingly started to embrace machine learning and its tremendous analytical and predictive power. As a result, a recent Morgan Stanley poll found that 51% of investment clients said machine learning was either a component or central to their investing process, up from 27% in 2016. However, some have found it difficult to harness the potential of big data and meld it with their traditional approach to investing.
Long road ahead
An immediate problem that investment companies face, especially smaller and medium-sized ones, is the fact that access to relevant talent in the machine learning space is limited. This means that building a suitable internal team that has knowledge of both machine learning and general asset management can be difficult and can take time. At the same time, finding a reliable third-party to incorporate machine learning can also be hard, as firms will need to make sure that any third-parties have produced measurable results in a timely fashion for previous clients.
Once the internal team is set up, the challenge then becomes developing the tools required to complete the implementation process. Consistently and rigorously applying machine learning to historical financial data is a difficult task; without the right tools at hand, the team can’t train and test the machine learning to the proper standard and results will ultimately suffer when it goes live.
Perfecting the process
Due to their complexity and pattern-recognition power, machine learning techniques are at a high risk of overfitting, which occurs when the technology learns to over-detect patterns that are present in the training data, yet disappear in a “live” environment. A strong implementation process can mitigate this, but the machine learning team will need to take a solid and rigorous approach to how the data is handled, as well as training and selecting which predictive models to use.
Communication is another crucial aspect of a good implementation process. The machine learning team will need to work closely with the portfolio managers to frame the problem that machine learning has been brought in to solve.
How the problem is framed has a significant impact on how the predictive models are applied in a real-life scenario – and thus what value they can bring to the business and its investment strategy. For example, if the problem is framed in a way that doesn’t fit the investment constraints of the fund manager, or the real trading conditions of the desk, the results, however accurate, will be mostly irrelevant.
Also, if the implementation process isn’t thorough enough, it will be difficult to account for factors that may not come into play until the machine learning process goes live. Commission structures, liquidity, and price delays can all impact predictive models during their live application in support investment strategies. This can mean that the machine learning models produced are accurate, yet fail to generate value when faced with the real market conditions they encounter in a “live” scenario.
Continuing together
For the marriage of machine learning and traditional methods to deliver the best results, there needs to be clear perimeters as well as transparency. This not only means being clear in what machine learning will be doing in terms of approach, but also what problems it will be solving during the investment process, such as supporting asset selection. This will make it much easier to see the benefits that the technology is bringing to the company.
Asset management firms should also leverage their machine learning partnerships for educational purposes, using methods like workshops to increase knowledge within their company and to provide greater transparency on what the machine learning tools are being used for. This approach can reduce some of the confusion which may exist within the team about where machine learning fits into their current processes and methods.
Moving forward
Implementing machine learning can be a complicated process with challenges at every step of the way. The talent pool is limited, suitable third-parties can be hard to find, and many in a company may be suspicious of the results it can produce. Additionally, a number of factors can impact the success of machine learning when it goes live, from overfitting to data leaks.
A strong implementation process, however, driven by good communication and a sound understanding of what the company wants the machine learning models to achieve, can mitigate these risks significantly. Ongoing communication between the machine learning team and rest of the organisation will also be essential, as it can prevent predictive models being siloed and underused by the company.
In order to achieve these goals, the role of machine learning must be clearly defined from the outset, and teams must be educated on how to maximise the potential of this technology. With these steps in place, machine learning can help active managers to evolve and re-gain their competitive edge.
