- Key insight: Haiqu developed a new technique to encode high-dimensional data (over 500 features) onto just 128 qubits, overcoming a major data processing hurdle for classical AI.
- What's at stake: For banks, even a small improvement in anomaly detection scores can significantly reduce the massive operational costs associated with investigating false positives.
- Forward look: While not claiming "quantum advantage" yet, Haiqu's CEO says the trial provides the "clearest empirical signal to date" that quantum-enabled machine learning "could soon be useful" for real-world data.
Overview bullets generated by AI with editorial review
Quantum software startup Haiqu announced results from a trial this week demonstrating that current quantum computers could detect subtle financial anomalies that could indicate fraud more efficiently than purely classical systems.
The research, which used a hybrid computing approach pairing quantum processing power with traditional machine learning models, revealed performance gains that suggest a near-term path toward achieving "quantum advantage" for large-scale, real-world problems.
Using conventional machine learning models with high-dimensional data, such as datasets with hundreds or thousands of so-called features that describe each transaction in a large set, can be expensive and challenging.
Additionally, fraudulent transactions are relatively rare (albeit potentially expensive) events, making up only a fraction of the total volume.
The core issue for traditional machine learning models is the computational complexity involved in processing this data, which can lead to difficulties in spotting the subtle, complex patterns indicative of sophisticated fraud schemes.
Haiqu sought to overcome this hurdle by developing a proprietary data encoding technique designed to prepare classical data for quantum processors. This technique allowed the firm to encode over 500 classical features onto 128 qubits.
"The ability to encode high dimensional data with hundreds and even thousands of features enables applications of a new scale, as what the team at Haiqu has experimentally shown on our hardware," said Jay Gambetta, director of IBM research, in a Friday press release from Haiqu.
Haiqu's CEO was somewhat circumspect about the advance.
"We are not claiming quantum advantage just yet," said Richard Givhan, CEO and co-founder of Haiqu. However, the company's trial provided "the clearest empirical signal to date" that "real-world, high-dimensional data can now be loaded onto a quantum computer" and that machine learning enabled by quantum computers "could soon be useful for processing this data."
What are features in financial data?
In the context of machine learning and financial data analysis, a feature (often used interchangeably with the terms "variable" and "attribute") is one single, measurable piece of quantifiable data used to describe an event, an entity or a financial transaction.
For banks and credit unions relying on data-driven models to identify suspicious activities, the ability to select and construct relevant features is a critical step for enhancing the performance of predictive models.
For each transaction, the features that a bank might analyze to try to detect fraud might be the transaction amount, the geographic location from which the transaction was submitted, the time of day it was submitted, the category of the merchant, the historical average spend of the payer, and the time since the last transaction.
In this example, the transaction would have six features. In Haiqu's trial, the data had 506 features.
It is often impractical to simply throw all of the available data about a transaction into the model. As such, banks are limited in the number of features they can choose to analyze for fraud, so they must be judicious about which ones they select.
However, Haiqu's trial introduces a new method to use a quantum computer to analyze more features at once.
The challenge of high-dimensional data
Modern fraud detection models typically analyze far more than six features; they incorporate hundreds and sometimes thousands of variables to build a comprehensive risk profile and catch subtle, complex fraud schemes.
A dataset with this many features is often called high-dimensional data, and this high dimensionality presents a significant challenge for classical computers running machine learning models.
As the number of features increases, the amount of data needed to reliably capture the unique patterns indicative of fraud grows exponentially, leading to what is often called the "curse of dimensionality."
Classical systems struggle to capture the intricate interactions among these features, which may be critical in detecting sophisticated fraud schemes.
Haiqu's claimed breakthrough lies in its ability to efficiently encode over 500 classical features onto a limited number of qubits (128 in their trial). By performing this complex process, known as quantum embedding, the quantum preprocessing step uses the unique properties of quantum mechanics — superposition and entanglement — to map the transaction data into a complex quantum state.
This quantum process aims to make the hidden, nonlinear relationships and correlations between these hundreds of features more visible, enabling the downstream classical classifier to achieve superior accuracy and speed in spotting the rare fraudulent anomalies that classical models might otherwise miss.
In essence, if the transaction data is a vast, tangled jungle of variables, the quantum preprocessing step acts as a powerful computational mapmaker, revealing the clear paths and hidden trails (i.e., correlations) that the classical systems need to follow to quickly identify the fraudulent activity.
"Haiqu's scalable embedding technology marks a turning point for quantum machine learning, making complex, high-dimensional data practical at scale," said Kristin Milchanowski, chief AI and data officer at BMO, a quote cited in a Friday press release from Haiqu.
Milchanowski added that this innovation "accelerates the shift toward real-world impact in industries like finance, where precision and insight redefine what's possible."
Key performance metrics
Haiqu compared its quantum-enhanced preprocessing technique against purely classical embedding methods using a critical metric for finance: the F1 score. (This score is not related to the motorsport competition.)
The F1 score is vital because it balances precision (avoiding false positives) and recall (catching true fraud events) on highly unbalanced datasets, such as transaction monitoring, where fraud is extremely rare.
The results, according to Haiqu, show a consistent trend favoring the quantum-enhanced approach.
In ideal simulation conditions, the quantum features achieved an F1 score of 0.98, significantly outperforming classical baselines ranging from 0.90 to 0.93.
Even when run on the real IBM Quantum Heron processor (which introduces noise, compared to the simulation), the model retained an F1 score of 0.96, demonstrating robustness under realistic device conditions.
"Anomaly detection is a very suitable target, since even a smaller improvement in scores can lead to crucial detections or elimination of false positives," noted Oleksandr Kyriienko, professor and chair in quantum technologies at the University of Sheffield, cited in the Friday press release from Haiqu.
Reducing false positives — alerts that needlessly burden fraud investigation teams — is a massive operational benefit for financial institutions due to the cost of these investigations.
Quantum edge for finance: Context and outlook
Haiqu's experiment builds on previous efforts by major financial players to validate the usefulness of quantum computation.
Earlier research demonstrated that quantum methods could improve financial applications. For instance, an experiment by HSBC and IBM showed a hybrid quantum-classical computing approach delivered up to 34% improvement over classical-only methods in predicting successful bond trades in opaque markets.
This kind of hybrid architecture — where classical systems handle data preparation and final decisions, and quantum algorithms handle the computationally intensive feature extraction or pattern recognition — is generally how experts anticipate banks will first integrate this technology.
"With Haiqu's software, quantum applications can run at a significantly larger scale," said Mykola Maksymenko, CTO and co-founder of Haiqu. "This is where the impact of quantum processing of data can become useful, as we see in our research on anomaly detection".
This research suggests that the ability of quantum mechanics to leverage superposition (where a qubit exists in multiple states simultaneously, allowing for parallel processing) and entanglement (where qubits are interconnected, facilitating complex pattern analysis) could soon provide the computational horsepower needed to keep pace with the increasing volume and sophistication of financial fraud.
