Case Study: Buy-Side Algo Credit Trading

Credit and Rates Trading Workflow Automation Best Execution Pricing and Liquidity Risk Analytics

Overview and Current Process

Buy-side asset managers are rapidly embracing artificial intelligence applications to price fixed income securities algorithmically in live trading execution environment or for purposes of end of day reconciliation and portfolio construction. The current fixed income capital market data flows are inefficient in many respects to enable robust coverage and precision for AI bond pricing. Markets remain heavily reliant on segregated and manual data operations between counterparties creating disparate data sets. These disparate data sets cause the market to suffer from information asymmetry and decentralization. As a result, insight from available data is fragmented and disseminated through manual exchanges between counterparties, which furthers the creation of disparate data sets.

Overbond performed detailed study of the current credit trading execution processes and how they could be improved with AI bond pricing model capability.

Example Of The Current Credit Trading Execution Process.

Two variations on how the buy-side trader initiates RFQ for a particular bond:

Electronic RFQ - Through a Bloomberg (or other) terminal window
1. Electronic RFQ windows appear on top of each other as per the time when the RFQ was initiated
2. The execution trader can confirm and submit the price suggested by the RFQ response, based on his/her knowledge of the bond market gathered during the day
3. The execution trader may also check a third-party pricing source to validate the pricing on the RFQ responses

Manual RFQ - Through a phone call
1. The security is checked on the Bloomberg terminal (or other) with the ISIN, as well as the “Quotes” on the terminal from other dealers, the best price to be accepted.
2. The phone call is treated as a priority and the electronically received RFQ responses are put on hold until a quote has been recorded.

Challenges With The Legacy Systems And Processes

When the RFQ is initiated by the execution trader via the Bloomberg (or other) terminal or a phone call, there are 4 main methods of validating the pricing responses for pre-trade best execution purposes:

CBBT taken from Bloomberg covering most of the cases
Fixed Yield – for the special type of the security
Price Range – a certain range is observed and validated
A certain spread range is confirmed – this might be quite wide

(depending on the RFQ responses received from electronic venues and/or over the phone)

The decision on which of these methods should be used to confirm best-execution, or whether the CBBT price is good would have to be taken by the execution trader on a case by case basis.

Execution Trader Workflow Objectives And Constraints:

The time window within which the execution trader is supposed to receive RFQ responses is 3 to 5 minutes; however, 1 to 2 minutes is considered market convention.
Difference between accepted best price and pre-trade suggested mid-price is measured as an indication of cost of execution, and/or difference between accepted price with market executed trade print just before and just after (on the tick level), needs to be between 5 to 10 cents on average.

Reasons driving the problem

The primary problem that the execution trader faces is due to the low confidence in prices suggested by Bloomberg CBBT and third-party applications that are currently used to validate best-execution pre-trade:

Bloomberg CBBT/ BGN / BMRK

The price suggested by Bloomberg CBBT (and or other Bloomberg pricing sources) is a composite price based on the most relevant executable quotations on Bloomberg’s Fixed Income Trading platform. The composite is based on current market activity for that bond. If the bond is illiquid, for whatsoever reason, the CBBT price would not represent the true price. This is especially the case when dealing with high yield bonds or bonds with varying liquidity profiles.

Current third-party pricing provider

The pricing methodology is based on a certain pre-defined methodology. The trader’s confidence in this system is low as the pricing methodology is not dynamic and this leads to the suggested pricing being incorrect. As perceived by the execution trader, there is a large difference in suggested price and accepted best price, 5-10 cents on average.

These factors lead to low confidence in the suggested prices and execution trading desks have hard time proving best execution to their clients based on the third-party impartial data feeds and pricing source.

Can AI Powered Bond Pricing be a solution?

AI Advantage over Statistical methods

COBI-Pricing, Overbond AI modeling techniques share many similarities with classic statistical modeling techniques starting from the fact that they both deal with volumus data. However, the key difference between statistical techniques and AI models Overbond applies is the goal of these approaches. While statisticians start with a set of known assumptions that are given to the model and best explain the expected behavior of the financial outcome in consideration, AI techniques rather aim at finding by themselves the method (with underlying assumptions that are unknown) that best predicts the outcome in consideration.

Clients can automate trading workflows using Overbond AI bond pricing feeds

Overbond COBI-Pricing, AI bond pricing feed can price bonds automatically with live refresh rates measured in sub 5 second range enabling automated trading workflow. Models can price 30% more bonds with low liquidity profile greatly increasing coverage and pre-trade best execution capabilities. Price is assigned for each bond under the issuer company and spread or yield curve is constructed. COBI-Pricing can systematically price a large number of liquid and illiquid securities, issuer names, and identify pricing tension metrics across large coverage book systematically.

COBI-Pricing was created as part of Overbond’s suite of AI algorithms for the fixed income capital markets. It algorithmically finds the most optimal best-executable secondary market bond price for global IG issuers utilizing machine-learning (ML) algorithms. The ML algorithms analyze millions of data points related to factors such as historical pricing trends between similar bonds and similar issuers, intra-day pricing volatility, trading volume and counterparty composition, company fundamentals trend, investor sentiment, and industry, rating or tenor cluster comparables. Data is aggregated from multiple types of data sources including:

Transactions data	Transactions occurring in the secondary market, and historical issuance spreads
Investment banking data	Fundamentals on corporations, their balance sheet indicators, proprietary data sets treasury groups of the corporations themselves had on file such as dealer quotations and trade points
Proprietary data	Client book data and/or direct access to large community of issuers and institutional investors via established feedback loops

Project Structure

Overbond structured the implementation of the buy-side best execution pre-trade pricing engine implementation project in two phases. Overbond first deployed and tested end of day data on a smaller universe of ISINs to algorithmically find the most optimal best executable secondary market bond price for each bond utilizing machine-learning (ML) algorithms. As mentioned before, Overbond ML algorithms analyze millions of data points related to factors such as historical pricing trends between similar bonds and similar issuers, intra-day pricing volatility, trading volume and counterparty composition, company fundamentals trend, investor sentiment, and industry, rating or tenor cluster comparables. Data is aggregated from multiple types of data sources and models are computationally intensive. That is why intra-day pricing was approached as a Phase 2 deliverable of the project.

Phase 1 Deliverables (Duration: 1 month)

Test web access to Overbond’s COBI pricing platform for a predefined list of 100 corporate issuers
Delivery of EOD Mid-Spreads as per the COBI Engine Output for agreed ISIN universe in the form of a table in a flat csv file
Delivery of calculated Bond prices based on the COBI output indicative of the mid price
Error/Difference – Analysis (COBI Price, BBG I-Spreads, Quoted Prices, Traded Prices)

Phase 2 Deliverables (Duration: 3 months)

AI Model Pricing

Front end interface developed to incorporate side by side visualization of the output of the finetuned and calibrated intra-day COBI pricing model

Intermediate results delivered with all ISINs/securities considered for intra-day pricing, confidence level of the modeled price and liquidity score for each ISIN
Input variables, such as aggregated transaction levels, fundamental data, rating and other

AI Model for TCA (Transaction Cost Analysis Post-Trade) and Best-Execution Pre-Trade

Bond Buyer Matching module output to serve as an input for the Best-Ex Optimization
Modelling trained on the trading book, executed trades in the past 2-year period
Results of AI model analyze the Distance-to-Cover for every RFQ initiated, ISIN/CUSIP pricing tension per counterparty, trading venue, and on top of the best executable price outputted by COBI-Pricing, model suggest best execution counterparty based on the past executed trades

How Overbond COBI-Pricing Algorithm works

The diagram below and the following paragraphs provide a description of how the Overbond COBI-Pricing algorithm works.

Data Intake & Pre-processing

The Overbond platform sources raw trading live date and fundamental data via Refinitiv’s live platforms. Our data sources include Refinitiv, S&P Global Market Intelligence (company level fundamental data) as well as various other sources. Overbond AI models have ability to incorporate dealer quotations/axes and investor preferences through direct feedback loops.

This raw data is then structured in the Overbond databases. Trading data and fundamental data are structured and mapped to the appropriate issuer ID. The data is systematically scrubbed for anomalies and null values. Finally, a set of key input factors are generated based on the raw input. These include but are not limited to factors that measure secondary market spread movements, recent issuance pricing levels, nearest neighbor credit ratings, and fundamental financial metrics. These factors are divided between sector and company-specific and are used as inputs to the machine learning models.

Model Training

The subsequent stage for the machine learning algorithm is to train and apply several models to calculate the output pricing levels. An Ensemble Learning strategy is used in three phases, meaning multiple models are combined to elevate overall robustness at each training stage. These models are each trained using a subset of the past data, ranging from one day, one month to a maximum of ten years. Advanced sampling techniques were used to account for data gaps for illiquid issuers in order to construct yield curves for all tenors and all issuers in coverage universe.

COBI-Pricing Data Intake

Successful data pre-processing is the key stage and pre-requisite for the COBI-Pricing algorithm operation. The precision of the algorithm output is critically dependent on the accuracy, timeliness, and relevance of the pre-processed input data. Overbond sources raw data from major data suppliers in the financial sector, including Refinitiv, Ice, The Six Group, EDI, MarketAxess, Tradeweb, Euroclear, Clearstream, DTCC, CDS, S&P Global Market Intelligence, major credit rating agencies, as well as other sources. The data COBIPricing algorithms uses includes the following:

Pre-processed Data	Source	Update Frequency	Relevance
Secondary market transactions data	Refinitiv, Ice, The Six Group, EDI, MarketAxess, Tradeweb	Intraday	Live prices and yields of companies’ bonds are used to measure spread movements in the coverage universe.
Settlement layer data per ISIN liquidity profile	Euroclear, Clearstream, DTCC, CDS	Intraday, end of day and historical	Settlement layer data when adjusted and merged with correct trading time stamp can augment the view into true liquidity of the particular ISIN as it contains settled trades by counterparties that were executed on the OTC level and otherwise do not appear in other consolidated data feeds.
Nearest Neighbour Credit per Rating and Industry Sector	Refinitiv, Ice, The Six Group, EDI, MarketAxess, Tradeweb, DBRS, Moody’s, Fitch, S&P	Weekly updates, quarterly filing cadence	Issuer’s industry sector cluster and past bond issuances and their ratings as well as composite rating for the issuer overall indicate the company’s risk level and benchmarking category. They are used to train the models and to backtest the accuracy of COBI-Pricing output.
Fundamental Financial Metrics	S&P Global Market Intelligence	Weekly updates, quarterly filing cadence	The company’s fundamental financial data is an indicator of the company’s credit-worthiness, and by extension, their cost of borrowing across tenors reflected in the bond pricing. In addition, fundamental metrics indicate the liquidity need of the company and its short term need to raise financing as well as leverage ratios. The financial profile of a company aids with clustering analysis of companies with similar characteristics.

COBI-Pricing Model Training for Different Liquidity Profiles

COBI-pricing is an advanced three-phase AI algorithm engineered to measure best-fit correlations with respect to company fundamental valuation and secondary market pricing for their bonds across sector peers and markets conditions at large. Models are tuned for different liquidity scenarios. A variety of preprocessed inputs flow into COBI-Pricing’s algorithm, to generate bond pricing output.

The first phase of the algorithm generates pricing curves for a list of companies specified by a domain expert with highest liquidity profile. This list contains companies from diversified sectors and are frequent issuers with liquid outstanding secondary market bonds (High Issuers). The Issuer should have a minimum number of bonds outstanding, bonds outstanding across the curve, and minimum number of trades and daily volume in secondary market for the algorithm to build a High Issuer curve using SVM algorithm.

The second phase uses a Nearest Neighbors algorithm to generate ISIN pricing curves for issuers with illiquid or insufficient secondaries (Low Issuers). Peers for each Low Issuer are identified using a score based on fundamental financial metrics, industry sector, credit ratings, secondary spreads, and issuances. The top High Issuer with the lowest blended score vis-à-vis a Low Issuer is classified as the peer. The secondary data from the top peer, along with the secondary data from the Low Issuer, is used to form an enhanced dataset for phase three to build curves.

The third phase creates pricing curves for all Low Issuers using Support Vector Regression models within SVM algorithm family, on the combined secondary data set of the Lower Issuer and the top peers as derived from the second phase.

COBI-Pricing handles the problem of sparse data sets, by filling the data gaps using credit matched peers with pricing levels to arrive at best fit or best executable prices for securities. Illiquid Companies with only minimal trading activity will now have modeled and relative-value prices for secondary market securities across all tenors. Their sparse data sets are enhanced with data from its peers, as determined in phase two of the algorithm.

COBI-Pricing Intra Day User Interface

COBI-Pricing AI output (data-feed) can be refreshed real-time or on an end of day basis depending on the user need. COBI-Pricing runs on the big-data set sourced from Overbond data lake. Overbond product team works with clients to customize coverage baskets based on trading style, models are then trained and back tested utilizing all data sources. The COBI-Pricing output can be integrated into a data feed via API, presented as custom visualization, or viewed on the Overbond Platform as downloadable table.

User Interface – Live Market Pricing:

Primary Market - Standard tenor live visualization
Secondary Market - ISIN pricing live visualization
Issuer curve live visualization

COBI-Pricing Intra Day Data Output

COBI-Pricing AI output via data-feed, API access, can be also refreshed real-time or on an end of day basis depending on the user need. COBI-Pricing runs on the big-data set sourced from Overbond data lake. Overbond product team works with clients to customize coverage baskets based on portfolio strategy or trading style, models are then trained and back tested utilizing all data sources. The COBIPricing output can be integrated into a data feed via API, presented as custom visualization, or viewed on the Overbond Platform as downloadable table.

Output Schema for Secondary Market Pricing

Trade Time	Issuer	ISIN	Benchmark (bps)	Spread (bps)	Yield (bps)	Price

Output Visualization for Designated Portfolio

Trade Time	Issuer	ISIN	Benchmark (bps)	Spread (bps)	Yield (bps)	Price
2019-09-02 08:38	Vonovia SE	DE000A18V146	-54.52	54.10	-0.42	109.66
2019-09-02 09:51	Deutsche Bahn Finance GMBH	XS0969368934	-55.91	17.44	-38.47	111.73
2019-09-02 11:57	Carrefour SA	FR0013342128	-56.03	47.73	-8.29	103.39
2019-09-02 12:39	Deutsche Telecome AG	XS2024716099	-4.34	100.35	96.01	105.71
2019-09-02 12:41	Vonovia SE	DE000A1ZY989	-50.14	63.80	13.66	107.57
2019-09-02 12:48	Iberdrola SA	XS1575444622	-51.50	32.10	-19.40	106.32
2019-09-02 13:39	LANXESS AG	XS0855167523	-56.89	40.81	-16.08	109.00
2019-09-02 14:07	Electricite de France SA	FR0011637586	-57.37	29.82	-27.54	104.19
2019-09-02 14:17	Danone SA	FR0013216892	-56.69	18.16	-38.53	100.60
2019-09-02 15:07	LANXESS AG	XS1501363425	-57.58	35.62	-21.96	100.87
2019-09-02 15:18	Siemens AG	XS2049616464	-57.72	20.71	-37.01	100.75
2019-09-02 15:47	Telenor ASA	XS2001737324	-47.70	37.99	-9.71	105.52
2019-09-02 15:51	Electricite de France SA	FR0011637586	-57.50	29.98	-27.52	104.19
2019-09-02 16:01	Vonovia SE	DE000A19B4D4	-57.89	39.93	-17.96	102.24

Backtest Approach

To establish that Overbond's COBI Bond Pricing algorithm deterministically constructs a fair value curve for a set of coverage issuers and can accurately price ISINs that are within the client’s coverage and with satisfactory precision and coverage. The set of coverage issuers and ISINs has been selected to represent diversified universe across issuers and ISINs with a liquid day-trading pattern, different ratings/risk profile, and bonds across curve. Back test description is per below.

Test Description
1. Measure precision of the fair value curve output from the algorithm by comparing algorithm output prediction for a specific issuer and tenor with actual secondary levels with same specific issuer and tenor bond
2. Test timeframe is all trades that happened in designated 2 time periods as defined (1st Sep 2019 – 15th Oct 2019) and includes a total 500-800 ISINs
3. Test actual traded levels on the day as a most-relevant comparison to modeled level performance by measuring the difference and comparing on a relative value basis difference with average bid-ask spread for that name and tenor (aggregating all trades on the day for the same ISIN/CUSIP)
4. Use reference data from secondary trades in individual bonds selected as a comparison value to the corresponding modeled algorithm price. Do not compare algorithm output to the data-set that was used as algorithm input to preserve integrity in all test cases.

Validation questions back-test should answer:

Does the model perform consistently and deterministically in building credit curves?
Is the accuracy of the model output validated by the fact that the difference between model individual ISIN price output and actual secondary trade is within an acceptable range (ie. 10 cents in price terms)?
Can the model perform well in pricing ISINs with very limited trading activity / liquidity profile?
Does the model construct accurate credit curves for issuers with few outstanding bonds or with illiquid bonds by using Nearest Neighbor AI approach?
Does the Nearest Neighbor algorithm ensure validity of relative value argument across sector and credit quality, effectively guaranteeing that best possible peers /comparables are utilized in pricing benchmarking?
Does the application of Support Vector Regressor (SVR) which uses non-linear regression to build credit curves, effectively ensure that no significant pricing curve distortions exist when longer or shorter tenor pricing is determined across the curve?

In order to test the yields suggested by COBI, back-test methodology has been defined with various metrics that would compare the prices/yields generated to various reference prices from legacy proprietary systems and/or Bloomberg data feeds. Prices suggested by COBI model should be within 10 cents of traded best price and ideally within traded best price and legacy system suggested price.

Back test visualization approach:

COBI-Pricing model provided spreads converted to price through the pricing library and stored them in a local database both with spread and price levels
The results are joined with various reference data elements including OMS reference executed trades prices, received RFQs record, Tradebook, Bloomberg CBBT to provide confidence in the way COBIPricing functions and can deliver results that can be used and trusted by the execution trader
The results are visualized to allow the stakeholders to interact and understand them

Metrics for error analysis:

Following metrics are used for prices and i-spreads against Tradebook actual trades and quotes, Refinitiv/Bloomberg, received RFQ records and OMS executed trades:

Mean Absolute Error
Mean Squared Error
% within Bid/Ask or Mid/Ask
Number of times trend break was predicted
Error where trend jumps or breaks occur
Performance in Time

Back Test Results

The primary goal of the back test is to compare prices generated by the pricing engine that uses COBI yields as input and compare them to:

Prices of the best received RFQ response
Automatic legacy system price (Bloomberg CBBT + Automatic Legacy Adjustment)

The expectation from trading execution desk is that prices based on COBI yields should effectively correct the automated system price:

i.e. Modeled price – Trader price < Automatic system price – Trader price

In order to visually compare prices, all prices are scaled using automatic system price and scored accordingly. As you can see in the sample graph below displaying results for 18 ISINs, in most cases COBI-Pricing level is within desired blue line range (BBG XYZ BID vs. BBG XYZ ASK), which is the category that indicates that COBI price is in line with traded price.

As we know, not all bonds fit perfectly on an issuer curve, some might be trading off the curve for a number of reasons, while employing a curve-builder approach to price ISINs in the secondary bond market we acknowledge that the model will not be precise for 100% for the entire bond universe. Hence, we segregated the bonds based on various measures including liquidity and volatility to identify bonds which can be priced accurately, and which bonds cannot be due to various market behavior, trading pattern and data availability limitations. COBI-Pricing model output price for each ISIN also has these explainability features attached to it:

Liquidity Score – indicates deep liquidity profile of the ISIN, based on bid-ask spread wideness, trade count and trade volume on the day and in recent history, peer ISIN comparisons, OTC flows from settlement layer data, price volatility intra-day and monitoring of the distribution of all corelated factors.
Confidence Score – measures confidence of the modeled COBI-Pricing output price for each ISIN is within a pre-defined threshold, ie. clients can set the threshold to be within 10 cents on the price basis
Confidence Interval – same as above but provides a range for trader to consider the high and low point

Business Impact

Over the past couple of years, we have witnessed profound changes in the fixed income marketplace with counterparties increasingly adopting quantitative investing and liquidity risk monitoring techniques to automate their trading workflows. These include systematic algorithmic trading and liquidity risk management automation, merging of fundamental discretionary and quantitative investment styles, consumption of increasing amounts of alternative data, and adoption of new methods of pricing analysis for fixed income instruments such as AI analytics like Overbond COBI-Pricing algorithm.

Specific use cases for the COBI-Pricing algorithm application are examined to identify business objectives and key benefits below. Overbond client organizations include buy-side institutions with significant trading volumes and with over $2+ trillion of combined AUM. Their innovation groups actively explore new technologies that can serve as the catalyst for trading execution automation, pre-trade best execution, post-trade fixed income TCA, and custom AI analytics.

AI Application	Business Objectives	Key Benefits
Intelligent automation and enhanced decision-making	Automate pre-trade execution, liquidity risk monitoring and post-trade TCA reporting Accurately measure pre-trade distance to modeled mid price as indication of cost of execution Significantly increase trade execution desk ability to measure and report on best execution Use auto-pricing models to improve reporting and input into risk systems	Overbond COBI- Pricing AI models aggregate historical data from multiple sources and optimize pricing for bonds with varying liquidity profiles. Overbond COBI Pricing can enhance users’ bond trading execution workflow by providing precise executable prices in up to 30% more situations when there is no directly observable trading price in the market. COBI-Pricing enables fully automated bond trading execution workflow with various curve visualizations and front-end trade analytics tools that are natively integrated with trader’s desktop, OMS or other internal desk system via API data feeds. As a buy-side trade execution desk, boost your best execution pretrade and post-trade TCA reporting with independent pricing feed, discover deep bond liquidity profile, and predict counterparty behaviour and where to route the RFQ request. Buy-side credit and rates trades can automate their execution and pricing capabilities with precision and confidence.

Implementation Considerations

Institutions considering AI predictive analytics implementation and big-data transformation projects, can employ acceleration utilizing externally calibrated models and market signals. Below are several key considerations and questions for executives in charge of AI roadmap:

What is the current state of our fixed-income in-house data?
What are our data science and engineering capabilities?
Are we building AI capabilities to grow revenue or cut cost?
How can we redefine the boundaries of our data universe or identify alternative data sources necessary to feed AI engine?
Given that AI learning curve is steep where do we begin?
How do we create and execute AI proof of concept use cases rapidly?
What are key success factors for our AI roadmap?

Custom AI Services
Overbond works with clients to identify and recommend practical AI analytics use cases that are aligned with strategic goals of the financial institution. We help assess current-state AI capabilities, and define roadmap to help clients realise value from AI applications. We manage cross-channel data flows across multiple systems and enable custom font-end visualizations.

Proven Methodology
With our targeted approach and implementation methodology, we quickly demonstrate value of AI analytics to test use cases, enabling client-side change management approach and stakeholder buy-in.

Operational Acceleration
We help clients build and deploy custom AI solutions to deliver proprietary analytics and tangible business outcomes. Our experience combines calibrated models, design patterns, engineering and data science best practices, that accelerate value and reduce implementation risk.

AI Analytics As-a-Service
Overbond helps customers design and oversee mechanisms to optimize and improve existing fixed income credit valuation, issuance and pricing prediction and pre-trade opportunity monitoring using AI. Our team of worldclass data scientists and engineers manage an iterative implementation approach from current state assessment to operational handover.

About Overbond

Overbond specializes in custom AI analytics development for clients implementing trade automation workflows, risk management, portfolio modeling and quantitative finance applications. Overbond supports financial institutions in the AI model development, implementation and validation stages as well as ongoing maintenance.

Contact:
Vuk Magdelinic | CEO
+1 (416) 559-7101
vuk.magdelinic@overbond.com