Training data is the historical sample used to build, tune, or validate a trading rule or model. It determines what the rule has actually learned, and a rule tuned repeatedly on one stretch of history risks being fitted to that stretch's noise rather than to something likely to recur.
How it works
A pattern-recognition rule encodes criteria, some of which may have been adjusted to perform well on a particular historical sample.
When that adjustment happens repeatedly against the same data, the rule can start describing the sample's quirks instead of general market behavior.
Why it matters
The sample decides whether a rule's behavior is real or accidental. Small samples, one market, or one regime produce numbers that may not transfer.
Testing on data the rule was not tuned against is how you check whether the rule describes the market or just the past.
A simple market example
A scanner rule is adjusted until it produces an impressive hit rate on last year's charts. That hit rate describes last year's data. To check it, the rule must be run on a different stretch of history it was never tuned against.
Common mistakes
Trusting a rule's performance without knowing what data it was built on.
Repeatedly tuning one rule against the same history, which is curve-fitting in another form.
Frequently asked questions
What makes training data good?
A large, representative sample covering different market conditions, and data the rule was not over-tuned against.
Can I trust a rule tested only on recent data?
Be cautious. Recent data is one regime; the rule should also be checked on other stretches.
Is training data the same as backtesting?
Training data is the sample used to build the rule. Backtesting is running it over historical data to evaluate it. They overlap but are not the same thing.
Educational content only. Definitions describe common market usage and may vary by jurisdiction, instrument, or institution.