What Is a Feature?
In machine learning, a feature is an input the model uses to make decisions.
Trading features can include:
- 20-day return
- Volume change
- Volatility
- Post-earnings gap
- News sentiment score
- Active on-chain addresses
The model does not “understand markets.” It searches statistical relationships in those inputs.
Why Features Matter More Than Models
Beginners often ask whether to use XGBoost, LSTM, or Transformers.
In trading, the bigger problem is usually the feature itself.
If features use future data, survivorship samples, or wrong timestamps, a powerful model simply learns wrong answers.
Three Common Feature Traps
1. Look-ahead bias
Using information that was not available at the trading time.
2. Time misalignment
Treating after-hours earnings data as if it was known during the session.
3. Too many features
If you test 500 indicators, some will look good by chance.
4 Questions for a Feature
- Was it available at the time?
- Was it updated fast enough?
- Does it have economic meaning?
- Does it work across different periods?
If not, a beautiful backtest is dangerous.
Quiz
Q1. A feature is the model’s:
A. Input B. Fee
Q2. Using future information is:
A. Look-ahead bias B. Normal optimization
Q3. A good feature should be:
A. Available at the time and economically meaningful B. Pretty in backtest only
Answer Key
Q1: A Q2: A Q3: A
Further reading: IBM — Machine Learning · Investopedia — Look-Ahead Bias
For education only. Historical feature performance does not guarantee future usefulness.
Get the pre-trade checklist.
We are turning these guides into a searchable checklist for checking terms, rules and risk before you trade.