A trading signal is a rule generated instruction to act in a market.
A trading signal is a rule generated instruction to act in a market. On its own, a signal is only an output. Its practical value depends on the process that created it, because the process determines whether the signal is a repeatable expression of a tested idea or a fragile artifact of noise.
In a systematic workflow, signal generation begins with an explicit hypothesis, formalized rules, and a defined research path from idea to publication. Sonar Sciences describes a research to publishing workflow in which research is turned into a structured artifact through specification, validation, and publication steps. That framing matters because it makes the signal traceable. A traceable signal can be linked back to the assumptions, transformations, tests, and controls that produced it. This is the basic mechanism by which process discipline affects signal value: the stronger the chain of evidence behind the signal, the more informative the signal is likely to be.
The first requirement is rule based construction. A signal derived from explicit rules can be reproduced, audited, and challenged. An ad hoc heuristic cannot be evaluated in the same way because the generation logic is not fully specified. Process discipline therefore starts before any performance metric is computed. It starts with whether another researcher could regenerate the same signal from the same inputs and obtain the same result.
The next requirement is protection against backtest overfitting. Sonar Sciences’ backtest overfitting audit tool is designed to examine whether a strategy or signal generation process has likely adapted too closely to historical noise. This matters because many apparently strong signals emerge only after repeated searching across parameter choices, feature combinations, filters, or time windows. In that setting, the observed backtest result may reflect selection bias rather than a durable relationship.
An overfitting audit changes the interpretation of a signal. Instead of asking only whether the signal looked good in sample, it asks whether the research process that found it was likely to have produced a false discovery. This directly supports the claim that signal value is determined by process robustness. If the process cannot survive an overfitting audit, then the signal’s apparent quality is not reliable evidence of underlying value. If the process is disciplined enough to reduce data mining and is then checked with out of sample evaluation, the signal has a more credible basis.
Out of sample testing is central here. A signal that retains its properties outside the data used to build it is more informative than one that only works in sample. The reason is mechanical. In sample performance can be improved by fitting to random quirks in the historical record. Out of sample performance tests whether the generation rules captured something that generalizes beyond the calibration set. A robust process therefore treats out of sample validation as part of signal creation, not as an optional add on.
The same logic appears in the deflated Sharpe ratio. Sonar Sciences’ glossary explains the deflated Sharpe ratio as a Sharpe ratio adjusted to account for multiple testing and non normal return effects. This is important for comparing signals generated by different research processes. A raw Sharpe ratio can look impressive even when it was produced after many trials. The deflated version asks a stricter question: after accounting for the number of trials and the statistical properties of the returns, how much evidence remains that the observed Sharpe is real rather than a lucky outcome.
This makes the deflated Sharpe ratio a process sensitive metric. Two signals can present similar raw performance statistics, yet the one produced by a broad, undisciplined search will generally be penalized more heavily than the one produced by a constrained, pre specified pipeline. The difference does not come from the signal output alone. It comes from how the output was obtained. That is the core reason the generation process determines value.
For research teams, this has a practical implication. Signal evaluation should not stop at the signal level. It should include an audit trail of hypothesis formation, data handling, parameter search breadth, validation design, and publication criteria. Sonar Sciences’ research to publishing workflow illustrates this as a chain rather than a single step. A signal moves from idea to formal research object through documented stages. Each stage either preserves or degrades confidence in the final instruction.
This also explains why publication quality and signal quality are linked. A published signal with documented construction rules, validation logic, and statistical context gives other researchers a basis for replication and critique. A signal shared without that process context is much harder to trust, even if its headline metric appears attractive. The absence of process evidence lowers informational value because the user cannot distinguish a robust rule from a lucky pattern.
The broad lesson is simple. A trading signal is not valuable because it exists, and not even because it performed well in one backtest. It is valuable only to the extent that the process behind it controlled for overfitting, defined its rules clearly, validated it out of sample, and measured its evidence with statistics that account for multiple testing. The signal is the last link in the chain. The chain is what determines whether the output deserves attention.
Covered in depth in the From research to publishing pillar hub.