AI trading bots
Machine learning is genuinely used in professional trading. The systems marketed to retail investors usually have little to do with that, and fail on the same fundamental problems as any other strategy.
Upfront: large trading firms genuinely use learning systems, and successfully. That's not marketing. Their systems just have little to do with what's offered to you online.
The difference is in the ingredients. Professional systems run on data sources that cost a lot of money, with millisecond execution and teams of specialists. The edge sits in those ingredients, not in the phrase artificial intelligence.
For any sold system, the question from the guru lesson still applies: why is someone selling it instead of using it themselves? A working system earns more with capital than with customers.
And one fundamental problem remains regardless of how good the technology is: a model learns from the past. The more degrees of freedom it has, the better it describes that past, and the greater the danger that it only describes that past.
Machine learning on financial-market data suffers from an unfavorable signal-to-noise ratio. The number of independent observations is limited, while the number of possible features is very large, which favors overfitting. Compounding this is non-stationarity: relationships change because market participants react to their exploitation.
The professional user's edge lies mostly outside the model: exclusive or expensive data sources, low-latency infrastructure, access to cheap execution, and the ability to bundle many weak signals across a large number of positions. These components aren't available to retail investors and can't be replaced by software.
For offered systems, the usual checks apply: a complete, externally verifiable results record over several years including costs, disclosure of the number of variants tested, stability against parameter changes, and a nameable economic cause. Missing any of these makes evaluation impossible.
Summary
- The professional edge lies in data and infrastructure, not the model.
- Financial data has little signal and a lot of noise, which favors overfitting.
- A sold system tells you something about its actual source of income.
Did you get it?
Why is machine learning on financial data especially difficult?
Little signal, a lot of noise, limited independent observations, and relationships that change over time.
Where does the professional user's edge lie?
In data, infrastructure, execution costs, and bundling many weak signals, not in the model itself.
What is non-stationarity?
Relationships change over time, partly because market participants react to their exploitation.
Related
- Gurus and signal groupsFraud Protection
- Copy tradingHype Check
- Who really moves the marketsStage 0