Can you predict which mutual funds will do better than others?
It turns out you can.
The number of investment products available for purchase increases every year. But wide product selection may create problems, both for the client and the advisor. The former can get lost in numerous options and refrain from taking any action. The latter may not advise properly – even having specific asset allocation in mind – because it is almost impossible to track every instrument which meets the initial client’s criteria. Overall, choosing the right product at the right time can make all the difference and determine success or failure.
At Comarch, we’ve decided to build an algorithm to address that issue.
What we were after
We wanted to provide necessary help for clients and advisors, assuming that some mutual funds can bring systematic returns in given market conditions. Depending on a fund’s strategy, methods of management and product specifics, it may be a better fit for purchase today, or remain more promising for the conditions of tomorrow – in a different environment.
In this vein, we decided to check whether AI-based algorithm can find the best mutual funds using interconnections hidden in product prices, macro indicators, market indexes, FX rates and commodities. The twist was – we did not try to decode these dependencies but rather put them to work and see the actual results in product ranking.
How we did that
We gathered macro indicators of major global economies (US, EU, China, Japan, LatAm, India), commodities prices, FX rates and major indexes quotations. Then, we employed price history for over 4000 mutual funds traded on NASDAQ. Our goal was to create a ranking of the funds and determine which ones would score the highest rate of return in the next quarter.
To create optimal environment for the algorithm, we needed to prepare the data we had accumulated and run a lot of additional analyses (e.g. return and risk ratios, fund efficiency indicators). Eventually, we have obtained thousands of market and product features as input data. The last thing to be done before the training begun was to effectively reduce data dimensions using Principal Component Analysis (PCA). Then, we proceeded to examine which type of AI model would be suitable for our kind of data series.
What carried the most promise
The most promising results came from the ensemble machine learning approach. It uses multiple learning algorithms (dozens best performing AI models in our case) to obtain a better predictive performance. Such performance usually cannot be obtained from any of the constituent learning algorithms alone. The models required months of tweaks and adjustments to produce the results we have today. And how we measured those results?
For the results measurement, we came up with a Top-Bottom Analysis. In the analyzed group, we created a ranking of mutual funds, sorted by the probability of achieving the highest rate of return in the next quarter. Then, we moved 3 months in the future and checked whether we were right, especially whether the top 10% was significantly better than the bottom 10% from our ranking.
To create the most realistic environment, we measured our results using so called test data, which is a dataset which the algorithm has never “seen” before. It guaranteed that such results can be obtainable in day-to-day operations.
In order to evaluate a mutual fund ability to outperform its peer groups, our model took various parameters into account: fund risk and past performance in different market conditions, return persistence and repeatability, risk-adjusted performance ratios as well as systematic risk related to main market factors. The last one, in particular, is somewhat vague: information-driven market movements which are quite unpredictable, human behavior which is uncertain, adaptive and sometimes repeatable, finally the fundamental economic laws that should explain dependencies between economic cycles and asset prices in a long run but not always do. We expected machine learning to tell us more where the truth was – comments Grzegorz Prosowicz, Consulting Director for Capital Markets at Comarch.
What the results were
The actual results for the test dataset went beyond our expectations. The best results we have achieved were 78% in Top-Bottom Analysis. It means that in 4 out of 5 cases we can predict which funds would score a higher rate of return.
Moreover, the difference in rate of return between top and bottom funds amounted to 2,99 percentage points quarterly (12,51 p.p. CAGR). In other words, advisor choosing products for their client from available product universe for one-year period, could count on 12,51 percentage points better returns with the algorithm – no matter how the markets perform.
Imagine having a service of that kind in your institution. Your advisors would finally get an effective help with choosing the right products and become more productive. Your clients would score higher ROIs and notice the competitive advantage you provide. Maybe even you could create model or robo-advisory portfolios automatically… Comarch AI-based Ranking Algorithm can have many faces - that is for sure.
Is the service going to reach a similar success in every country or circumstances? We are going to find it out and make it even better – that is what we do. Is the service ready to launch as cloud API in your institution or get tested by you in form of proof of concept? Definitely.
Dominik Łyżwa, Business Consultant at Comarch