Since November 2015, our machine learning tool for detecting spoofing and market manipulation has been getting a lot of buzz. The company has a record of detecting market manipulation in listed derivatives based upon high profile cases. We put together an Artificial Intelligence (AI) tool that allowed compliance officers to quickly detect who was manipulating the markets and why.

But computers don't really trade - they are programmed, just like our trade surveillance tool. Surely these computer programs looking at trade data wouldn’t be able to catch a sophisticated scheme. After all, the public cases against Navinder Sarao and Michael Coscia didn’t show any really deep levels of sophistication.

My plan was to show that I was better than the machine. In doing so, we could use my trading to improve Neurensic’s machine learning trade surveillance tool. To keep this above reproach, this scheme would be implemented in a fictitious market, the Jelly Bean Exchange, and would trade a fictitious product.  

The trading scheme I chose to go after is called “collapsing.”  Here is the premise:

Step 1: I put in a series of small orders at different prices. Similar to layering but my orders would be small, so they wouldn’t stand out. I had no intention of these orders ever being filled.

Step 2: I put in my real order on the other side of the market.

Step 3: I collapsed the small spoof orders to the same price point to create a spoof size. On the market data stream, it would look like big size had just hit the book at the same price point. In reality, it was just a bunch of orders all being collapsed together.

Step 4: My real order was filled. This is the order I really wanted filled.

Step 5: I cancelled all my small spoof orders.

This test was included in a series of other trading activity when we turned it over to our programs to evaluate. There were plenty of other things for the system to catch and I figured it would miss my “collapse” strategy.

So how did I do?

The Neurensic compliance engine scores all “clusters” of trades on a scale from 200 to 800. A low score (200) means that the engine is saying that the cluster of trades was very much safe, and not similar to known cases of market manipulation. A high score indicates that there is a strong similarity in the trade cluster to known cases of market manipulation.

My collapsing strategy was scored at 752. They caught me. Artificial intelligence really does beat my ability to manipulate markets. I’m glad I was only trading Jelly Beans inside our own office.  

Maybe I should buy our technologists some Jelly Beans for a job well done.