>Nasdaq runs more than 40 different algorithms, using about 35,000 parameters, to look for market abuse and manipulation in real time.
I laughed when I read this. 40 whole algorithms!
I laughed when I read this. 40 whole algorithms!
Click link:
> Nasdaq North America Surveillance team monitors 3 equity markets, 6 options markets and one futures market with real-time surveillance and post-trade surveillance of unusual market activity. The surveillance department is monitoring the markets for Insider Trading, Fraud and Manipulation, including manipulation through trading—pump and dump—and order book manipulation—spoofing and layering, as well as handling events in the market such as clearly erroneous transactions.
> The surveillance program today is using algorithmic coding to detect unusual market behavior running over 40 different algorithms in real-time, looking for market abuse and manipulation. The patterns have sophisticated algorithms that use approximately 35,000 parameters. In addition to real-time surveillance, there are over 150 patterns covering post trade surveillance, which are used to identify a wide range of potential misconduct.
> The activity is monitored across equity and options markets, with some market-specific alerts and some alerts encompassing data from all markets.
So they have 40 different predictive models running in production (a lot!), and 150 hand-written (or single-pattern) rules. They also employ active learning with experts and this reduces their training data needs with 95%. 40 models using 35000 features, is 875 features per model on average. All this bickering is typical.
Why pick on the 40 algorithms?
I think this definition will be important in law because it is where we will define responsibility for actions of autonomous things.
That's how you write honestly for the target audience.
> Nasdaq runs more than 40 different algorithms, using about 35,000 parameters, to look for market abuse and manipulation in real time.
That is how you use big words on an audience not capable of seeing through their meaninglessness, and it's the responsibility of people who do see through it to call it out.
As someone else mentioned. The 35k parameters is skeptical. Taleb and Tversky and Kahneman have good evidence that most algorithms are better with less parameters. The more parameters, the more noise.
Interesting observation.
Do you have more specific references? Those three are so popular it’s hard to narrow to commentary on parameters and noise.
Basically, as your signals approach infinity, your chance of finding correlation approaches 1. Taleb argues your chance of finding causation is likely to decrease.