AI trading bots
Machine learning really is used in professional trading. The systems offered to private investors usually have little to do with it and fail on the same basic problems as any other strategy.
First of all: big trading firms really do use learning systems, and successfully. That's not marketing. But these systems have little to do with what's offered to you online for $99 a month.
The difference lies in the ingredients. Professional systems work with data that costs a lot of money, execute orders in milliseconds and are looked after by whole teams. The edge lies in these ingredients, not in the words artificial intelligence.
For a system being sold, the question from Automated Trading also applies: why is someone selling it instead of using it themselves? A system that works earns more with its own money than with customers.
And one basic problem remains, however good the technology. A model learns from the past. The more knobs it has, the better it describes that past, and the greater the danger that it only describes that.
Machine learning is especially hard on financial data, because the real signal is tiny and the noise huge. There are only a limited number of independent observations but a great many possible features, and that invites overfitting. Also, the relationships change. As soon as many people exploit a pattern, it disappears.
The professionals' edge lies mostly outside the model: expensive or exclusive data, technology with extremely short paths to the exchange, cheap execution and the ability to bundle many weak signals across a very large number of positions. Private investors have none of this, and no software can replace it.
With systems on offer, the usual checks help: is there a complete record of all results over several years, including costs, that someone independent can check? Is it stated how many variants were tested? Does the result hold if you change the settings a little? And is there an economic reason why it should work? If any of these is missing, you can't judge the system.
Summary
- The professionals' edge lies in data and technology, not in the model.
- Financial data have little signal and a lot of noise, which invites overfitting.
- That a system is being sold reveals where its provider's money comes from.
Did you get it?
Why is machine learning on financial data especially difficult?
Little signal, lots of noise, few independent observations and relationships that change.
Where does the professional user's edge lie?
In data, technology, cheap execution and bundling many weak signals, not in the model itself.
What is non-stationarity?
That relationships change over time, partly because market participants exploit them and they disappear as a result.
Related
- Gurus and signal groupsFraud Protection
- Copy tradingHype Check
- Who really moves the marketsStage 1