When a bond is unavailable for an asset manager to buy, perhaps due to liquidity constraints, the manager may search for a substitute bond with similar features. Methods exist to help the manager find a replacement, such as a machine learning approach known as random forest.
Now, two hedge fund managers, Dario Villani and Kharen Musaelian, present another way based on a technique called quantum cognition machine learning, or QCML.
The method, developed in conjunction with a team from BlackRock and published in Risk.net earlier this month, is best suited to high-yield bonds.
“In investment grade bonds, random forest and QCML give similar results,” Villani explains, but “in high yield bonds or municipal bonds, where the amount of data is very limited and there is a lot of sparsity…and outliers, we do way better than any random forest”.
Villani and Musaelian are theoretical physicists by training, who co-founded machine learning firm Qognitive in 2023.
For the past two years, they have been finessing their QCML method, applying it to problems in finance as well as in other areas such as medicine.
In this episode of Quantcast, Villani and Musaelian explain that their approach overcomes the curse of dimensionality that afflicts classical machine learning methods. “If you have a large number of features, the data requirements go up exponentially,” says Musaelian, and because of that, “humans figure out which features to focus on and…block out the rest of the data”.
Quantum mechanics tackles uncertainty differently, without discarding information or features that classical methods would consider irrelevant. What is key to the method is its ability to reduce the effective number of dimensions.
This property allows QCML to handle a large number of features and train a machine learning model even with a low number of observations for each feature. Villani and Musaelian’s method uses quantum mathematics but currently runs on classical computers.
Another application was developed in collaboration with Deutsche Bank and BlackRock quants to solve a similar problem in equity. There, QCML is used to identify linkages between firms that are economically connected, as explained in a separate working paper.
But applications go beyond finance. One case study the duo has been experimenting with is the analysis of data related to genomic instability of cancer cells. The dataset included 140 features of about 200 patients, a dimensionality classical methods struggle with. “A priori, nobody knows which ones are important and whether you can actually make any predictions,” says Musaelian. “We do well on that set, we beat all the classical methods precisely because the model can blur out whatever is not relevant.”
In a much more extreme example, Musaelian says the model manages to extract some meaning from a dataset of 16,000 features related to only 88 patients. “Noise washes out all the classical methods that we have compared our method with, whereas QCML can be robust to very large amounts of noise,” says Musaelian.
Another possible application of QCML is to detect similarity among claims to an insurance company. That is an example of a high-dimensional problem, only some of whose features are relevant, but not known beforehand.
Despite contributing to the advancement of AI and its applications, Villani is sceptical of doomsday-like predictions about the technology supplanting the role of humans. A current debate on the matter has had prominent names including Nvidia’s Jensen Huang and Meta’s Yann LeCun opining against the views of others, such as the executives of Anthropic, an artificial intelligence developer, who argue that AI may replace half of junior office roles within five years.
For Villani, at the base of these opinions is the idea that language models achieve intelligence – which he disputes. “Language is not intelligence and actually can get in the way of intelligence of reasoning,” he says, alluding to today’s tendency to look at language models as a tool to develop creative solutions, while they are only a representation of pre-existing knowledge. “In statistical learning, at the end of the day you have a loss function and a bunch of gradients,” he says.
The conversation concludes with a focus on the joint development agreement with IBM to explore the use of QCML on quantum hardware. “Recently we did an experiment on the Heron machine and we keep making progress,” Villani says, explaining that the focus is on problems that require more than 50 qubits, which cannot be solved on classical hardware.
Index
00:00:00 Introduction
00:02:15 Beating the curse of dimensionality
00:11:10 Core principles of the quantum approach
00:24:43 Mapping spaces and quantum fidelity
00:34:20 Similarity detection in bonds and equity
00:42:13 QCML applications medicine and robotics
00:51:52 The impact of unregulated AI adoption on humanity debate
01:01:17 QCML on a quantum hardware
01:05:17 The beginning and future of the approach
To hear the full interview, listen in the player above, or download. Future podcasts in our Quantcast series will be uploaded to Risk.net. You can also visit the main page here to access all tracks, or go to Spotify, Amazon Music or the iTunes store to listen and subscribe.