Market Abuse and Advanced Analytics
Do Predictive Models play a role in Market Abuse Solutions?
By John Doherty - Senior Presales Consultant, SAS Australia
10 years ago when I was new to Market Abuse surveillance solutions we used a Ferrari to catch Ferrari’s. That is a low latency, high throughput algo engine to create rules, listen for patterns and publish alerts when the Market Abuse rules were triggered. In the past few years I started fielding questions about Predictive Analytics and Machine Learning and that discussion continues today.
I won’t claim to be an expert in machine learning, but now that I’ve moved to a company whose pedigree is Advanced Analytics I’ve learnt what predictive and machine learning models are, how at a high level they are built and I feel more comfortable in opening a discussion on their place in Market Abuse surveillance.
If you train a model against the Market Abuse your system identifies, your models recreate the rules you already have in place. It all becomes self-fulfilling unless sufficient instances of unidentified fraud is reported as with credit cards.
Harnessing Predictive Models for Fraud Detection
Neural Network, Linear Regression, Random Forrest, Gradient Boost are but a few types of Predictive Models. The models are created by mining and statistically analysing historical data to determine the probability of an event when other identifiable events occur. The models identify the most predictive variables and how they should be parameterised to recognise the abuse going forward.
Predictive Models, particularly Neural Networks are great for many, many fraud types including credit card, online and other digital transactions to identify and stop fraud attempts that discrete rules don’t find. These models often give fraud detection systems uplifts of 50% or more and stop crime in its tracks. The data to create the models is available because when credit card fraud occurs and customers report it, meaning the way the fraud was perpetrated is captured in the fraud data used to train models to defend against it.
When fraudsters’ change their behaviours and model accuracies start to deteriorate, (but we know it because the amount of fraud reported increases) he models can be effectively retrained with the newly reported fraud data to uncover the new behaviours, “unknowns.”
Sometimes Market Abuse does get reported (e.g. when the banker who was colluding with his bureau of statistics friend is reported by his broker for being too lucky too often around retail figure announcements) it typically goes back to the regulator and isn’t included in the data an organisation can reference to build a model. Furthermore, who really knows how much Market Abuse occurs? The ‘fingerprints’ of a crime having taken place, and the feedback loop (like what exists with credit card fraud when money disappears and customers report it) is simply not in place.
The Market Abuse that trading operations’ identify is typically based on surveillance systems triggering discrete rules or by analysts finding a needle in the haystack when manually trolling through immense volumes of trade and order data. If you train a model against the Market Abuse your rules detect, your models can identify predictive variables that you were not aware of and non-predictive variables you currently use. So you should be able to optimise your rules’ performance with respect to true and false positives but it’s unlikely that you’ll get the uplift that you do against credit card and other types of fraud when “system unidentified” instances of the fraud are included in the data used to train the models.
So can Predictive models play a role in Market Abuse systems?
Yes, if you have the data required to train one. A well trained model should work just as well in a Market Abuse situation as it does in other fraud and compliance surveillance solutions.
If you don’t have the data what can you do?
Regardless of whether an organisation has the data to build predictive models; accurate, efficient and holistic Trade and Trader Surveillance systems require a balanced and hybrid approach to meet their regulatory obligations. A hybrid approach for a holistic solution would include a number of approaches and tools including;
|Primary Use Cases
Identifying known Patterns of behaviour
|Anomaly detection &
|Tell me when something (person, trade volumes etc.)
behaves differently than their peers or historically
|Unstructured Text Analytics
|Tell me when these words are contained in a phone call, message, email.
|Show me who’s connected to who and how they are connected
|Predictive Models /Machine Learning
|Uncover unknowns and optimise rules
Who might have the data required to build accurate Predictive models to detect Market Abuse?
The Regulators. Regulators have aggregate Market Abuse data across participants, markets, asset classes and abuse types. Regulators obtain reported cases and their associated data from what should be all sell side organisations using different abuse detection systems and rule sets (both vendor and in-house) and could perform cross rule comparisons. Regulators see Market Abuse across the full spectrum of market participants from day traders, to buy side institutions and corporates, to sell side brokers and banks. Regulators receive tips and collect data/evidence on abuse cases that systems don’t pick up.
Regulators are one of the most likely candidates for the data required to build accurate predictive models that identify Market Abuse that goes unrecognised and unrecorded by rule based systems today.
Should regulators build and maintain Trade & Trader Surveillance models to be deployed within their spheres of influence? That’s a discussion for another day.