Currently, the only solution in the marketplace that shows what bonds have traded and at what prices across various trading venues that also includes the Over-The-Counter (OTC) trading volumes, is “client side data aggregation”. This involves aggregating data feeds from various providers internally. Up until a few years ago, it wasn’t possible to process, map or normalize all these different data feeds in real-time. Today, AI analytics makes this possible.
A fragmented market leads to fragmented data
Both sell side dealers and buy-side asset managers are increasingly relying on AI applications to price fixed income securities in live trading and automate part of their daily workflows. For example, Greenwich Associates report published in September 2020, estimates approximately 45% of the European bond market is currently traded electronically, compared to 38% a year ago. However, this volume is spread across multiple competing trading venues, and most data sources specific to one trading venue do not have enough coverage to provide traders with a view of true liquidity across the market and consequently price precision that can be executed automatically.
The bond market heavily relies on segregated data disseminated between counterparties. This creates fragmented data sets coming from different composite data aggregators, trading venues but also does not cover the bulk of OTC traded flows amounting on average to 60% to 70% of total traded daily volumes. For example, many vendors compete for market share, such as Bloomberg FIT, Tradeweb, MarketAxess, MTS. Large OTC data flows that trade through a dealer network by phone contribute to further fragmentation of course. In addition, FINRA Trace and other OTC regulatory mandated trade reporting repositories outside of the US dollar denominated market, suffer from reporting lags. FINRA Trace in particular, has a built-in 15 minute lag, which can sometimes be double that time in actuality accounting for processing and computation delays.
Harnessing data properly from multiple unique sources is the critical step before deriving insights from it, notes Simon Steward from Capital Group. “Traders have to have access to vast amounts of data that both validate and challenge outcomes,” he adds.
Automated trading requires client-side data aggregation
Overbond’s engineering and data science teams are applying multi-level AI analytics to address these problems:
1. The delays in incoming data feeds, and;
2. The sheer amount of data that needs to be mapped and reconciled to the reference master and then processed in real-time in order to automate the live trading workflows.
Overbond’s team set a several second refresh rate as the maximum target speed threshold for AI bond pricing, liquidity measurement and the trade risk algorithm refresh rate. This is a very hard problem to solve, given that most of the single data source models in the market refresh with a time lag of several minutes. To aggregate data across venues and different data types, the time lag often increases 10 fold.
The need and demand is there though. According to Charles Elkan, global head of machine learning at Goldman Sachs, the financial services industry seems like fertile ground for AI to generate profit.
Advanced parallel processing in cloud architecture
Overbond is the first fixed income AI analytics provider that was able to successfully apply advanced parallel processing in serverless cloud architecture. We utilized AWS Kinesis and Lambda solutions to break down all mapping and data processing steps and apply heavy parallelization.
Overbond’s algorithms collate and organize large volumes of disparate data, including non-traditional data sets such as fundamental and settlement layer data. Using novel AI liquidity scoring, AI models can tier all trades. Trade tiers then determine if the trades qualify for full-automation, trader supervision, or should not be traded at the current time.
The main difference comparing Overbond’s approach and other bond pricing models in the marketplace is that our engine aggregates data from multiple client-side sources into a single tenant container. This is what we refer to as data aggregation “client-side”. The engine not only produces fair value pricing optimization in real-time, it also computes risk metrics such as liquidity and confidence scores that segregate all trades into Tiers. These Tiers are one, two, three, and/or not recommended right now.
Enabling trade-automation real-time
This tiering approach of trades can enable trade-automation. In effect it is like performing trading analysis for each ISIN at each time it trades. A Tier one trade (single ISIN at the point of time during the open trading day) would have the most stable and high liquidity profile in the last 24 hours (or 8 hours of trading). It would have relatively large trading volume and relatively lower pricing volatility in this same time horizon. As a result, we could expect to see lower than 7 cents pricing precision (on average and across a very large number of ISINs tested safely over multiple years of market data).
So naturally, Tier 1 trades would be a candidate for fully automated execution, with no trader involved, or in other words, without touch execution. Tier 2 trades correspond to a higher risk. They have a lower pricing precision – under 8 cents – on some of the large back tests Overbond’s data science team has performed. Depending on the trading desk, these are usually candidates for one-touch execution. The trader would need to supervise Tier 2 trades and trade execution would happen with just one click.
The Overbond platform sources raw live trading and fundamental data from a range of suppliers. These include Refinitiv data composites, S&P Global Market Intelligence, ICE, EDI, Euroclear and FINRA TRACE as well as major credit agencies. It also collates company-level fundamental data, dealer quotations, internal client executed trade records and investor preferences through feedback.
Why is client-side data aggregation real-time important?
And why does the fixed income market require a refresh rate of a few seconds to automate trading workflow? Simply put, without client-side data aggregation, trading workflow cannot be automated in a meaningful way. For example, on any given day in the U.S. corporate bond market, roughly 70% of the trades are executed for 100 bonds or fewer. Those 100 bonds also traded on average on 2 to 3 different electronic venues, so available pricing data coverage is fragmented even further. Considering that average size corporate trading desk has 1,500-2500 bonds in their coverage universe, and we can get latest pricing for only 2-4% of those bonds, poses major limiting factor when it comes to trading automation.
Automating is also easier said than done. It is not only collating the aggregate data or mapping it that is necessarily hard. The technological challenge is performing data aggregation, mapping and reconciliation, then historical modeling and AI optimizations necessary all in only a few seconds. This is where the application of serverless cloud technologies such as AWS Kinesis Analytics with Lambda functions parallelization of processes proved transformational. AI algorithms that would normally take hours to finish calculations are now processed in seconds.
1. “Data in Fixed Income Trading”, written by Jon Williams, Refinitiv, published by Markets Media on 18.06.2020 [https://www.marketsmedia.com/data-in-fixed-income-trading/]
2. “Future-Proofing’ the Trading Desk”, published by Markets Media on 22.06.2020 [https://www.marketsmedia.com/future-proofing-the-trading-desk/]
3. “Will Machine Learning Transform Finance?”, video conversation with Charles Elkan, Managing Director, Goldman Sachs published by Yale Insights on 03.01.2020. [https://insights.som.yale.edu/insights/will-machine-learning-transform-finance]
4. “The Challenge of Trading Corporate Bonds Electronically”, report published by Kevin McPartland from Greenwich Associated on 13.05. 2019 [https://www.greenwich.com/blog/challenge-trading-corporate-bonds-electronically]
With offices in Toronto, Montreal and New York, Overbond is an artificial intelligence (AI) quantitative analytics provider for institutional fixed income capital markets. We provide data aggregation solutions and a comprehensive suite of AI algorithms (known as COBI) for bond pricing, liquidity risk, auto-execution and hedging, pre-trade signals and market surveillance. Founded in 2015, Overbond is transforming how global investment banks, institutional investors, corporations and governments connect and access bond markets’ AI data analytics.
Our fully-digital platform and suite of AI models utilizes the world’s most advanced algorithms in finance. We produce market analytics with AI pricing and trade-tiering optimizations that enable full trade automation. Our tools and processes facilitate faster, more precise trade execution, eliminate information flow inefficiencies, reduce execution costs and minimize compliance risks. In addition, the Overbond platform assists market participants with systematic price discovery and liquidity risk management, enabling sell-side traders to execute more profitable trades in larger volumes and buy-side traders to achieve pre-trade best execution and counterparty selection.
Our global client base comprises buy-side institutions with over $2 trillion of combined AUM and corporate and government issuers with more than $20 billion in bonds outstanding. These clients include the European Central Bank, the European Stability Mechanism (ESM), Wells Fargo, Fidelity, ATB Financial, Hydro One and EPCOR. In addition to its three offices, Overbond has R&D labs with three leading Canadian universities and the Institute for Data Valorization (IVADO).
For more information about Overbond and using AI analytics to automate trading please visit www.overbond.com