top of page

Katana Labs IP

Background

Developed in-house at ING, Katana was a project devised to assist portfolio managers and traders with day-to-day decision making. Driven by machine learning technology, Katana’s adaptive approach to bond market trends utilizes big data to provide high quality positioning insights to professional traders and portfolio managers. The digital advisor was developed in collaboration with some of the largest asset managers in Europe from the outset.

Having proven effective during the preliminary stages, the pre-trade analytics tool advanced from prototype to market-ready in just over a year, and in light of its success, in 2019, ING decided to launch Katana as a standalone business.

A platform for data-driven analytics, Katana seeks to track pairs of bonds and make swap recommendations to profit from market dislocations. Katana monitors the relative credit spread for traditionally paired bonds using big data and machine learning to provide insight into bonds’ relative value by tracking market behavior.

 

When a divergence emerges in bonds that have historically seen paired trends, Katana will advise that a swap is made to the cheaper offering to capitalize on the correction from the depressed bond price. Katana is expected to predict market compression with around 90% accuracy.

Katana used similarity factors such as currency, country of risk, credit rating, and sector to identify the pairing of bonds to provide a single similarity score. The bonds are given a present z-score to describe their current disparity, compared to historical data for these pairings. An algorithm is then used to predict the likelihood of compression in the bond pairs spread. The platform also includes a powerful set of filters to help users navigate the opportunity space efficiently.

The model can deliver an extensive overview of the most significant opportunities from screening nearly 200-million bond pairings while providing relative value insights and trade ideas for bonds from North America, Europe and Emerging Markets. The model monitors holdings and watchlists alongside broader market trends that allow asset managers and traders to identify opportunities more efficiently and confidently. There are two main processes involved in the model, one that identifies relevant pairs and secondly, one that identifies dislocations and predicts compressions.

Dislocations are shown to have been 90% accurate over a measured time of 30 days.

The technology is already integrated into the Bloomberg app portal with c. 330,000 users and its own web-based delivery platform providing serverless access to users.

Problem Katana Addressed For The Market

The ability to discover actionable trade opportunities that have a high probability of mean reversion in the bond market is a challenge because of the sparse distribution of bonds available.

Traditional proprietary trade discovery models (usually run on excel spreadsheets) are inadequate for discovering actionable trade opportunities because of factors such as bond selection, availability, and price appetite.

Katana's deep neural net is capable of time-series forecasting for 200 million fixed income market pairs.

Katana aimed to make portfolio discovery more efficient through the use of machine learning. The Katana model optimized the asset discovery process through rationality-based algorithms that looked at all possible bond pair opportunities that had a high probability of mean reversion and filtered them down to a user's specific criteria.

The model allowed asset managers to find anomalies and mispricings that were not always obvious in a large investment universe.

Whitepaper, Example Ideas and Neural Net Development

bottom of page