A trading signal is a structured recommendation derived from market data: enter long, exit, reduce exposure, or stay flat. Signal generation sits between raw feeds and risk-controlled execution. Its quality depends as much on definitions, filters, and timing as on the sophistication of the underlying formula.
From raw data to features
Raw ticks and candles are rarely used directly. Feature engineering transforms them into returns, volatility estimates, momentum scores, spread ratios, and cross-asset relationships.
Lookback windows define how much history each feature sees. Short windows react quickly but amplify noise; long windows smooth noise but lag structural shifts.
Stationarity assumptions matter. Features built on price levels behave differently from those built on log returns or z-scores. Document which transformation your signal expects.
Feature pipelines should be deterministic and versioned. A silent change in how a moving average is computed can alter live behaviour without any strategy parameter update.
- Returns and momentum — rate of change over defined horizons
- Volatility — rolling standard deviation or range estimators
- Relative value — spreads, ratios, z-scores versus baseline
- Microstructure — book imbalance, trade intensity, spread width
Signal types and semantics
Binary signals say yes or no: trade or do not trade. Continuous signals output a score or probability that position sizing logic later converts into notional.
Directional signals specify long or short bias. Neutral signals explicitly recommend no new exposure, which is different from missing data or system error.
Entry and exit signals can use different rules. Asymmetric design—tight exits, loose entries—is common but should be documented to avoid confusion in attribution.
Composite signals aggregate sub-signals with weights or voting. Conflicts between components should resolve through a defined rule, not implicit last-writer wins behaviour.
Filters and regime gates
Not every computed signal should reach execution. Liquidity filters block trades when spreads are wide or depth is thin. Volatility filters reduce size when ranges expand beyond historical norms.
Time-of-day and session filters respect market open dynamics, funding windows, and scheduled announcements. Automated systems that ignore calendars trade into predictable noise.
Correlation filters prevent stacking redundant exposure across related pairs. Without them, multiple signals may express the same underlying bet.
Regime gates switch signal families on or off based on classified market state. A mean-reversion signal during a strong trend can produce repeated losses until the gate flips.
Timing, persistence, and decay
Signals can be event-driven on each tick or bar close, or interval-based on a clock. Mixing bar-close logic with tick data without alignment creates lookahead bias in research and jitter in live trading.
Persistence rules require conditions to hold for N consecutive observations before firing. They reduce whipsaw at the cost of later entries.
Signal decay reduces confidence or size as time passes without reinforcement. Old signals should not retain full weight indefinitely unless the thesis explicitly requires it.
Cooldown periods block re-entry shortly after exit to prevent churn in sideways markets. Tune cooldown relative to typical holding period and fee structure.
Validation without overfitting
Count how many effective parameters and filters your signal stack introduces. Large search spaces on limited history produce signals that fit noise.
Hold out multiple non-overlapping periods and stress windows with different volatility character. Consistency across regimes is more informative than a single spectacular backtest year.
Measure signal turnover and implied transaction costs. A high-Sharpe signal that trades every bar may disappear after realistic fees and slippage.
Log signal values at decision time in live trading. Comparing distributions live versus research catches drift in data or unintended code changes early.
- Parameter discipline — limit search space versus history length
- Regime holdouts — test across unlike market periods
- Cost-adjusted turnover — net signal value after fees
- Live distribution checks — compare signal stats to research
Signals are contracts between data and execution: precise definitions, explicit filters, and logged values at decision time. Without that discipline, even elegant formulas become untestable noise in production.