Research & signal quality
How to evaluate trading signals beyond win rate
Assess trading signal quality using net expectancy, drawdown, sample size, execution costs, and transparent records instead of headline accuracy claims.
By atradeaday · Published
A useful trading signal is specific enough to test: instrument, direction, timestamp, entry condition, invalidation, target logic, and expiry. A vague bullish message posted after a rally is not equivalent to a timestamped plan that could have been followed beforehand.
Start with an auditable record
Request the complete signal history, including losses, cancellations, expired ideas, and corrections. Distinguish hypothetical, paper-traded, and live executed results. Screenshots of selected winners cannot establish a track record.
A high win rate can be compatible with losses if occasional losing trades are much larger than the winners. It can also be inflated by counting multiple targets from one setup as separate winning signals while counting the stop only once.
Use a net-expectancy example
Suppose 100 hypothetical trades contain 60 wins averaging 0.5R and 40 losses averaging 1R. Gross expectancy is 0.6 × 0.5 - 0.4 × 1 = -0.1R per trade, despite a 60% win rate. Costs make the result worse if they were not already included.
Conversely, a lower win rate can coexist with positive expectancy when average wins sufficiently exceed average losses. Neither case guarantees future outcomes. The distribution, sample quality, and consistency of execution matter.
Review a balanced scorecard
- Net expectancy: average realized outcome after the relevant costs.
- Drawdown: peak-to-trough equity decline, including open exposure under a stated valuation rule.
- Sample size and period: enough context to see how much evidence actually exists.
- Coverage: which markets, sessions, and conditions were included or excluded.
- Execution: how signal prices differ from achievable fills.
- Concentration: whether a few trades or one regime produced most of the result.
Keep definitions consistent across providers. Profit factor, for example, compares gross profits with the absolute total of losses under a stated accounting basis; it becomes unstable with very few losses. Ratios without their underlying trade distribution can mislead.
A confidence score is not an observed win probability
A model may assign a score because several features agree. Calling that score 80% does not establish that eight out of ten comparable trades will win. That claim requires calibration against later outcomes using a clear success definition and unseen data.
atradeaday provides technical and AI-assisted research. Its methodology guide describes the weighted technical score and its limitations; this library does not present an independently audited performance record.
For broader background, see Investor.gov on fees and expenses. Use backtesting controls before interpreting attractive historical numbers.