The claim by HSBC that noise in an IBM quantum computer helped deliver a 34% improvement in algorithmic trading performance has been disputed by experts, with Scott Aaronson, a leading academic in the field, saying that the finding had “more red flags than a Peoples Liberation Army parade”.
Published to considerable fanfare on 24 September, HSBC’s announcement was lauded by the financial press as a ‘Sputnik moment’. When US markets opened on the same day, IBM’s stock price rose by 5%.
The claimed breakthrough hinges upon the real-time estimation of fill probabilities of requests for quotes (RFQs) from a panel of dealers, which due to the complexity of datasets is best handled by machine learning. While HSBC focused on the algorithmic trading of corporate bonds, the problem is also relevant to equities in areas like block trading, over-the-counter derivatives and exchange-traded funds (ETFs) where RFQs are common.
Standard machine learning algorithms work by identifying features from a training dataset in order to make predictions with test data. Working in conjunction with IBM, the HSBC team used a quantum computer to generate new machine learning features using a superposition of quantum states. Back-testing this output with an area-under-the-curve (AUC) metric, the quantum approach improved on non-quantum approaches by 34%, HSBC claimed.
Yet HSBC conceded in its paper that the observed effect was purely empirical, caused by ‘inherent noise’ in the process with no theoretical foundation. That means the much-touted result is most likely a cherry-picked example of selection bias, according to Aaronson, who is chair of computer science and head of the Quantum Information Centre at the University of Texas.
Speaking to Global Trading, Aaronson said that quantum computing advances over the past 30 years were limited to two areas. “Simulating quantum systems, which might help for example with designing new drugs and materials and industrial processes, and 2) breaking most of the public-key cryptography (such as RSA) that currently protects the Internet”.
“Eventually, there should be some modest benefits for optimization and machine learning problems, including in finance, for example using Grover’s algorithm”, Aaronson explained, with the caveat that speeding up such calculations involved significant hurdles which limited their likely benefits to the distant future. “This is one way we know that people claiming to get such speedups today are almost always lying”.
Describing the HSBC paper as “scientifically risible” and “fatally flawed”, Aaronson added, “the ‘advantage’ is just a strange artifact of the particular methods that they decided to compare—that it has nothing really to do with quantum mechanics in general, or with quantum computational speedup in particular”.
An explanation of why HSBC might have stumbled on algorithmic improvement by accident comes from Morten Hagen, a data science expert who built machine learning models for American Express before launching his own company, Context Technologies. Hagen told Global Trading that a well-known method known as “random resampling with shrinkage” could potentially replicate the noise effect of IBM’s quantum computer. Shrinkage has a long history in the quantitative finance community, particularly in the estimation of covariance matrices.
“This means that you create a much larger dataset, built on the original data. In each ‘draw’ or ‘instance’ you replace some of the feature values with a new ‘sampled’ value that is ‘shrunk’ towards zero”, Hagen said. “Much the same as the quantum ‘black-box’, that would have the effect of normalizing the classical distributions.”
