How to Separate a Market Signal From a Trading Decision

A market signal begins as a rule applied to information. It might identify an unusual volume reading, a price relationship, or a change in a company measure. Before discussing what to do, state exactly what the rule observes, which inputs it uses, and when those inputs become available. A label such as strong signal is not a definition, and a bright indicator does not explain its own limitations.

This article introduces an original research handoff framework, not a validated trading strategy. Bailey and coauthors' original paper examines backtest overfitting, illustrating why historical selection can produce misleading confidence. That research does not validate the example rule or workflow here. The narrower aim is to prevent a detected pattern from skipping the additional reasoning required to consider any real-world action.

Primary-source context: Bailey, Borwein, Lopez de Prado and Zhu: The Probability of Backtest Overfitting.

Document the instrument universe, observation window, comparison method, and handling of missing data. If a rule compares volume with a historical average, say whether that average includes the current observation and whether the sessions are comparable. If inputs are incomplete, label the output incomplete. A number calculated successfully can still represent the wrong question.

Then write the interpretation as a hypothesis. Unusual activity could be consistent with several explanations; the rule alone may not distinguish them. Record which additional evidence would help and which explanations remain outside your ability to assess. Keep an explicit distinction between detecting a condition and predicting an outcome. A condition can be measured reliably without having useful predictive power, and an interesting historical relationship can fail when the data or market environment changes.

Create a handoff note with the observation, data-quality status, proposed explanation, contrary evidence, and unresolved questions. Only after that note is complete should the workflow reach a separate assessment of whether any action is justified. That assessment has its own questions about objectives, existing exposure, costs, liquidity, and implementation constraints. The signal cannot answer those questions merely by having a high numerical value.

Make no action an explicit possible result at every stage. Missing data may stop the analysis before interpretation. Weak evidence may stop it before a decision assessment. A coherent hypothesis may still lead nowhere because the proposed action does not fit the relevant constraints. These outcomes are not malfunctions of the signal. They show that the observation and the decision have different jobs and different standards of evidence.

Suppose a fictional rule flags a company when full-session volume exceeds twice the average of the previous ten complete sessions. In an invented dataset, the average is 100,000 shares and the latest complete session contains 250,000. The ratio is 2.5, so the condition is met. This arithmetic only establishes the defined relationship in the supplied data.

The first review discovers that the historical values cover one venue while the latest value covers several venues. The condition is therefore not comparable, and the handoff stops at data review. After obtaining consistent coverage, imagine the ratio is still 2.2. The observation now survives that specific check, but its meaning remains uncertain.

The research note lists possible explanations and seeks an issuer disclosure relevant to the original question. It does not label the event institutional buying or forecast a higher price. Even if a plausible explanation emerges, a trading decision would require a separate assessment. The example demonstrates a sequence of questions, not a profitable volume rule.

Define the exact rule, universe, timeframe, input availability, and missing-data behavior. Confirm that comparisons use consistent units and coverage. Save the calculation version and the underlying observation date so a later review can reproduce the signal that was actually seen, rather than a corrected version created afterward.

Write what the observation establishes and what interpretation you are adding. Include at least one competing explanation. Identify whether supporting evidence is independent or simply repeats the same source. If historical testing is cited, ask what was tested, how alternatives were selected, and what assumptions limit the result.

At the handoff, state whether the next step is data repair, further research, a separate decision assessment, or no action. Do not allow a default button or a colorful label to decide which of those outcomes applies.

A reproducible signal definition needs more than a formula. Record the observation timestamp, source availability timestamp, eligible instrument set, session boundaries, and calculation version. For the fictional volume rule, specify that the comparison uses the previous ten completed sessions and excludes the current session from the average. If the last ten sessions are not all present, the example rule should return insufficient data rather than silently using a shorter window and retaining the same label.

Consider a hypothetical event where the latest session finishes before the vendor publishes its final consolidated volume. A provisional number observed shortly after the close is not the same input as a later finalized figure. Save which version generated the signal. If the later number changes the result, create a revised observation rather than pretending the original calculation used information that had not arrived. This preserves the boundary between what the rule could observe then and what is known afterward.

The same boundary applies to any simulated action. A rule requiring a final session value cannot assume a fill earlier in that session on the basis of that final value. The specification must choose an eligible later execution point if it models execution at all, and explain the associated assumptions. That choice does not establish a realistic trading system by itself. It simply prevents a timing contradiction from being hidden inside an otherwise correct formula.

Use the original hypothetical data: ten previous complete sessions average 100,000 shares, so their combined volume is 1,000,000. The new complete session contains 250,000 shares. Under the specified rule, the ratio is 250,000 divided by 100,000, or 2.5. If someone instead averages all eleven sessions including the latest, the mean becomes approximately 113,636 and the ratio becomes 2.2. Both calculations can be performed correctly, but they implement different rules.

Near a threshold, that difference can change classification. Suppose another invented latest session has volume of 210,000 against the same prior total. Excluding the latest session gives 2.1. Including it gives an eleven session mean of 110,000 and a ratio of about 1.91. A condition requiring a ratio strictly above two would fire under the first definition and not the second. The disagreement is methodological, not necessarily a calculation error or evidence that one source is unreliable.

The handoff should therefore carry the denominator definition with the result. Also state how equality is handled: greater than two and at least two are different conditions at exactly two. These details matter for reproducibility, but they do not imply that either boundary has predictive value. The example threshold is stipulated for explanation. Refining its arithmetic does not justify selecting a slightly different threshold merely because that version would have produced a more appealing historical set of signals.

Imagine a hypothetical signal that survives the definition and coverage checks. The record says: ratio 2.2 using ten prior comparable sessions; final input available after the session; data version saved. The hypothesis field proposes that a new company disclosure may explain the activity. The researcher then finds a disclosure but cannot establish whether it accounts for the observed trading. The appropriate interpretation remains a possible connection, not a demonstrated causal explanation.

Next, imagine the researcher considers a hypothetical implementation exercise and discovers that the relevant cost inputs are unavailable. That blocks a credible net outcome simulation even if the observation itself remains valid. The final status can be data valid, explanation unresolved, implementation unevaluated, no action. Those separate states are more useful than a single rejected label, because a later update might resolve the implementation assumptions without changing the original signal calculation or settling its interpretation.

Preserve each stage as a short dated entry. If the missing cost information later becomes available, reopen the appropriate stage rather than rerunning the history with newer data and erasing the original limitation. This is particularly useful when multiple people or tools handle the workflow. A downstream reviewer should be able to see which claim has actually been checked. Passing the input gate means the defined observation is reproducible, not that the explanation is established or that a transaction has become justified.

A signal can be discussed without simulating a trade, but if an example includes one, write the execution assumptions openly. Suppose a purely hypothetical simulation buys 100 shares at 20 currency units and later sells them at 20.30. Before costs, the price difference contributes 30 units. Assume a fee of 5 at entry and 5 at exit, plus adverse execution of 0.03 per share on each side relative to those reference prices. The assumed additional execution cost is 6 units.

Under those stipulated assumptions, the remaining gain is 14 units: 30 minus 10 minus 6. Relative to the 2,000 unit reference purchase value, that is 0.7%, before any other omitted items. The 1.5% reference price move and the 0.7% simplified net result describe different calculations. Neither is a forecast, evidence of a profitable signal, or a complete account of all possible costs. The example merely shows why an observation about price cannot substitute for implementation arithmetic.

Specify whether reference prices already incorporate any assumed spread or execution difference so those costs are not counted twice. If actual fill assumptions cannot be supported, label the exercise conditional rather than calling it an executable result. Also keep an unfilled order distinct from a zero return completed trade. The missing execution outcome is a limitation of the simulated path, not a number to fill automatically because the signal engine successfully produced an observation.

Document how the rule was chosen before describing any historical results. In an original hypothetical research project, someone examines twenty volume thresholds, several observation windows, and two instrument groups, then selects the combination that looks best. The selected history reflects that search. A clear research record includes the alternatives examined, not just the final formula. Without that record, another reader cannot tell whether the rule was specified in advance or chosen after comparing many outcomes.

One practical proposed worksheet separates development from later evaluation. In the development area, record changes and why they were tried. In the evaluation area, record the version fixed before inspecting the designated later observations. If the later results prompt another modification, label the revised rule as a new development version. Do not continue describing those already inspected observations as untouched evidence for the revised rule. This is a bookkeeping boundary around what informed selection, not a guarantee that the remaining evaluation is sufficient.

Include unsuccessful and undefined results under a consistent policy. Missing inputs, unavailable execution assumptions, and excluded instruments should remain visible in the research record. A historical chart can look clean because difficult observations disappeared during preparation. The worksheet should explain each exclusion and whether the rule could have identified the eligible set at the time. These questions constrain what a historical exercise can honestly claim without requiring a forecast or suggesting that a particular test establishes readiness for live trading.

A reusable handoff can begin with four sentences: the rule observed blank; the inputs passed these specific checks; the proposed explanation is blank; and the unresolved limitation is blank. Follow with the permitted next research step. Avoid language such as confirmed opportunity when only the arithmetic has been confirmed. A label should say exactly what was checked, because later readers may see the label without the surrounding technical detail.

Consider overlapping detections in a hypothetical three day period. Each uses a window sharing most of the same sessions. Store three observation records if that is what the rule generated, but do not describe them as three independent confirmations of a prediction. Likewise, two indicators derived from the same underlying inputs need an explanation of what distinct information, if any, the second adds. Agreement is an observation to analyze rather than a shortcut around the hypothesis stage.

Finish the worksheet with an explicit unresolved state for each dependency. A missing input should not default to false, because false suggests the condition was evaluated and not met. A valid signal should not default to an order, because no implementation assessment has yet been recorded. These distinctions make the interface and written record match the actual research process. The useful output is an observation whose definition, availability, limitations, and next step can be inspected without inferring more certainty than the evidence supports.

A detected signal is not a probability of profit. Repeated detection does not create independent evidence when the observations overlap or share the same inputs. Agreement among several indicators may reflect the same underlying price series rather than different reasons to believe an outcome is likely. Likewise, confirming that a calculation is correct does not confirm the interpretation attached to it.

The workflow does not establish that signals are useless or that discretionary judgment is superior. It simply requires each stage to make a claim proportionate to its evidence. A research tool can identify something worth examining without recommending a transaction. Keeping that distinction visible makes the final record easier to evaluate and helps prevent the precision of a calculation from being mistaken for certainty about an investment decision.