Connecting infrastructure behaviors with business outcomes with machine learning analytics was the topic of CEO Donal Byrne's presentation at Intelligent Trading Summit.
I had a chance to hear our CEO, Donal Byrne, discuss the changes happening in trade infrastructure analytics at Intelligent Trading Summit NYC this week. One of the key points he made was that the next level of analysis involves connecting the dots between business execution results and what the underlying infrastructure is actually doing.
He explained that a gap exists today between understanding business outcomes and the infrastructure impact on those outcomes, despite the intrinsic linkage between them. The gap persists because there are people and analytics focused on how the infrastructure is performing, while a different group of people and analytics are focused on trading outcomes.
To that end, Donal’s presentation outlined some of our ongoing efforts examining which aspects of AI and machine learning are useful and reliable for addressing that gap for specific use cases, for example:
It was interesting to see how often other participants and panelists seemed to echo similar views throughout the event. For example, the #tradingsummitnyc twitter stream quoted Anne Petersen, Global Head of Sales at Pico saying "There is not a competitive advantage to have access to networks. Understanding how your environment will operate in various conditions is how you will create that advantage."
Another panelist noted that: “Analytics are about understanding what the algos and plant are doing and correlating that to what the market is doing. The advantage comes making this analysis an intraday activity so you can change trading behaviors.”
Clearly, the interest is there. In his presentation, Donal explains that our experience has shown is that there is a hidden step 1a that many projects can trip over while trying to address this interest.
With analytics there are often two goals. The first is to use the results to get a clear understanding of what is happening, ideally allowing you to see connections and correlations that were not possible before. The second goal is to leverage that new understanding to further examine and explore the results to obtain insight into the most effective ways to change the process or transaction to improve future results.
However, before you can have a clear understanding of what is happening, you must have confidence in the results being produced by the analytics. In other words, the hidden step 1a involves demonstrating that the analysis results are of high enough quality that they do deliver an accurate understanding.
This lesson we learned all too well deploying our streaming analytics to help trading businesses understand where and how to optimize end-to-end transaction performance. Every unexpected answer was traced back to the specific transaction, through every infrastructure hop, and down to the messages and packets themselves to prove the accuracy of the intelligence being produced by our streaming analytics. The benefit of this experience is that trading firms now have a high-quality source for operational and business outcome intelligence for their AI and machine learning efforts.
This same process holds when applying machine learning analytics to. For example, we asked an “off-the-shelf” machine learning algorithm to learn what the factors influence the success/fail outcomes of IOC orders from the operational and business outcome intelligence being produced by our streaming analytics. Then to determine the accuracy of the results, use them to predict the success/fail outcomes of a different set of IOC orders. Those initial results were not good – it accurately predicted the outcomes less than 50% of the time. In other words, a coin flip had a better chance of being right than a “learning bot” downloaded from the internet.
It is only by blending multiple learning algorithms, leveraging the strengths of each, did the accuracy improve to +90% levels. Only then can there the confidence in the understanding (i.e., goal 1) that Order Type was a top 3 factor affecting the success/fail outcome of IOC orders.
Only then can there be confidence in the insight (i.e., goal 2) that for strategies dependent on the “Limit Orders” type, then investing to optimize latency across that order-lifecycle is a must for success. Meanwhile, for strategies dependent on a “Pegged Order” type, then it might be time to take a closer look at what understanding and insights can be gleaned from the market data and pricing operations.
So the key to successfully making machine learning part of intraday activities is doing that hidden step 1a work to develop reliable analytical algorithms and then deploying it on a platform capable of keeping up with the volume and velocity intraday trading operations.