Who this is for — Readers who want to distinguish the information produced by a model from the position to be taken and the order that will reach the market.
A systematic trading signal is the output of a rule or model that summarises information available at a given time to support a decision. It can express direction, rank, probability, expected return, risk or state. It is not yet a position and it is not an order.
The distinction is essential. A positive signal can lead to a zero position because a limit is active; several signals may be combined before sizing; a desired position may generate only the order needed to close the difference from the current holding. Confusing these layers makes both the backtest and error attribution opaque.
Six objects not to confuse
| Object | Meaning | Example unit |
|---|---|---|
| Feature | transformed information available at decision time | return, spread, standardised surprise |
| Forecast | estimate of a future or conditional quantity | expected return, volatility, probability |
| Signal / score | summary used by the decision rule | sign, percentile, continuous score |
| Target position | desired exposure after constraints and sizing | currency amount, contracts, weight, delta |
| Order | instruction to move from the current to the target position | quantity, type, limit, validity |
| Fill | actual execution reported by the venue or broker | quantity, price, timestamp, commission |
The relationship need not be one-to-one. A continuous forecast may become a signal only beyond a band; the portfolio optimiser may attenuate correlated signals; one order may produce several fills or none. Preserving intermediate states makes it possible to determine whether a P&L deviation originates in the model, sizing or execution.
Signal types
- Binary or categorical: long/flat/short or other discrete states. It is easy to interpret, but the threshold can discard information and must be validated.
- Continuous: measures intensity or conviction on a scale. It requires a normalisation rule and a mapping to risk.
- Ranking: orders instruments at the same point in time. The result depends on the available universe and treatment of missing values.
- Probabilistic: estimates a conditional probability. Discrimination alone is insufficient: calibration, the base rate and the cost function affect the decision.
- Return or risk forecast: retains economic units, but must state the horizon, currency and reference distribution.
- State signal: enables, limits or disables another rule; examples include data quality, regime or operational risk.
A signal may describe alpha or risk. Reducing exposure when estimated volatility rises is a systematic decision, but it is not the same as predicting the sign of the return.
Timestamps and availability
Each signal value requires at least:
- the economic or market period to which its inputs refer;
- the time at which each input actually becomes available;
- the time at which preprocessing and the model can finish;
- the signal publication time;
- the horizon to which the signal refers and its expiry.
A fundamental referring to March but published in May cannot enter an April signal. Normalisation fitted on the whole dataset transfers information from the future even if each individual feature appears lagged. A ranking built from today's index constituents can exclude former constituents in advance. The signal therefore inherits the contract of point-in-time data.
From signal to position
An explicit mapping answers four questions:
- how scale and sign become desired exposure;
- how signals on the same instrument are combined;
- how correlation, concentration, leverage and liquidity are handled;
- what inertia or band prevents economically wasteful turnover.
For a simple score sᵢ,t, an illustrative notation is:
raw weightᵢ,t = f(sᵢ,t); target weight = constraints(normalise(raw weight))The function f and the constraint operator belong to the strategy; they are
not neutral details. Clipping, ranking, volatility scaling and neutralisation
can dominate the risk profile. The target position must also be compared with
the current position: the order concerns the difference, possibly rounded and
limited by capacity.
State, missing values and revisions
A signal should not assume implicitly that every input exists. The
specification distinguishes at least valid, stale, missing, invalid and
suppressed. Filling a missing value with zero may mean “no forecast”, “neutral
forecast” or “the observed value is actually zero”: these interpretations are
different.
When a model is recalibrated, the signal version changes. Correcting historical data after the fact may be useful for analysis, but it must not silently rewrite what the production system observed. Both the “as observed” and corrected series are useful for audit and attribution.
Illustrative example
Suppose a cross-sectional score ranges from 0 to 100. The strategy forms positions only in the ranking tails, reduces weights under a sector constraint and does not trade when data are stale. One instrument moves from a score of 52 to 91, but the concentration limit prevents an increase in its position: the signal is strong, the target remains unchanged, and no order is created.
The example shows why “the signal said buy” is an incomplete description. It does not suggest universal percentiles, thresholds or rules.
How to evaluate it
| Level | Check |
|---|---|
| Informational | relationship with the forecast object, uncertainty and out-of-sample stability |
| Economic | incremental value relative to a baseline and costs |
| Portfolio | contribution after combination, constraints and common exposures |
| Operational | latency, coverage, stale rate, revisions and errors |
| Governance | version, owner, use limits and deactivation procedure |
Accuracy or correlation alone is not enough. A rare but economically important forecast, an uncalibrated probability or a high-turnover signal requires metrics coherent with the decision. Evaluation must respect temporal order, multiple testing and costs.
Sources
- U.S. Securities and Exchange Commission, Staff Report on Algorithmic Trading in U.S. Capital Markets — distinction between algorithmic decision and execution functions and their risks.
- Joseph Simonian, CFA Institute Research Foundation, Investment Model Validation: A Guide for Practitioners — validation, benchmarks, monitoring and documentation for investment models.
- Tilmann Gneiting and Adrian E. Raftery, Strictly Proper Scoring Rules, Prediction, and Estimation — principles for evaluating probabilistic forecasts without separating the score from the quantity forecast.
- Narasimhan Jegadeesh and Sheridan Titman, Returns to Buying Winners and Selling Losers — academic example of a cross-sectional ranking signal under a specific design.