Selection bias · baseline · rejection

6M Indicators: Benchmark Them or Choose None

By Kyle ParrottUpdated August 13, 2026

Five indicator families, ten lookbacks, four thresholds, two bar sizes, and two exit rules create 800 combinations before the first chart is judged. The best-looking line in that tournament can be pure selection noise. “Best” is therefore not an indicator name; it is a comparison claim that has to survive a declared baseline, later data, costs, and an honest possibility that nothing wins.

Count the tournament Use the decision card

AcceptStable task-specific improvement after costsResearch onlyInteresting but not holdout-qualifiedChoose noneNo candidate beats the baseline

Hidden degrees of freedom

The Word “Best” Creates a Statistical Tournament

If a researcher tries enough correlated transformations, some will look unusually good in the sample by chance. The family includes every indicator, lookback, threshold, direction, bar construction, filter, exit, and discarded variation actually inspected—not just the final line shown.

candidate family = inputs × parameters × rulestasks × horizons × filters

Why the count matters

NIST’s multiple-comparison guidance explains why repeating pairwise comparisons does not generally control the comparison family. In indicator research, undisclosed trial count makes an apparently exceptional outcome hard to distinguish from selection luck.

Evidence boundary: this page has not run that tournament on 6M. It reports no winning indicator, lookback, threshold, accuracy, return, drawdown, Sharpe ratio, or fill result.

One transformation, one job

Give Every Indicator a Declared Decision Task

An indicator cannot be “good for 6M” in the abstract. A transformation may be evaluated as a state descriptor, eligibility filter, sizing input, trigger, or exit. Mixing these jobs lets a weak entry rule inherit credit from a useful risk normalizer or vice versa.

Describe state

Classify past-only trend, volatility, participation, or distance from a reference without making a directional trade claim.

Filter eligibility

Permit or block a separately defined setup. Compare the setup with and without the filter under identical risk.

Scale exposure

Map a risk estimate to a bounded size rule. Judge forecast calibration and tail loss, not raw direction accuracy.

Trigger or exit

Define the precise state transition, observable timestamp, order instruction, invalidation, and expiration.

A code-ready task statement

“At the close of each eligible 30-minute bar, test whether a past-only volatility-percentile filter improves the net risk-adjusted outcome of a separately frozen breakout rule versus that same rule without the filter.” The indicator is not allowed to change the entry, exit, or size unless that is its declared task.

Instrument: Dated 6M
Decision: One function
Comparator: Same rule, no filter
Finding: Not reported

Specification before inspection

Register Candidates as Equations and Timestamps

Names such as RSI, moving average, ATR, or volume profile are incomplete. Freeze the source field, contract, bar clock, formula, initialization, lookback, threshold, comparison operator, signal availability, missing-data behavior, and expiration.

Family	Possible measurable task	Required controls	Common invalid claim
Trend	Past-price slope or relative position used as a filter	Past-only window, warm-up, contract roll, level adjustment	“Moving averages work on 6M”
Momentum	Signed change or oscillator state used for a declared trigger	Threshold family, overlap, clock match, reversal definition	“Oversold means price must rise”
Volatility	Range or return dispersion used for eligibility or size	Estimator, sampling interval, event state, tail treatment	“Low volatility predicts direction”
Participation	Completed volume, trade rate, spread, or depth used as a quality gate	Feed fields, dated expiry, missing books, session and size	“High volume confirms the buyer”
Composite	Frozen combination tested only after components are benchmarked	Interaction registry, added complexity, incremental ablation	“More confirmations make a signal stronger”

CME’s Market by Order FAQ distinguishes order-level, price-level, and consolidated depth information. A volume or order-book indicator can use only what the selected feed actually records.

Complexity must earn admission

Make the Candidate Beat a Naive Alternative

A positive net result is not enough. Persistent market direction, variable exposure, or favorable event concentration can make many transformations appear useful. The baseline should perform the same decision job with less information and share every execution, cost, sizing, and eligibility assumption.

Directional task

Compare with no directional information

Use no-trade, constant direction, or a simple past-return rule as appropriate. Match gross exposure, holding time, session, and risk so the indicator cannot win by taking more risk.

Filter task

Ablate the filter

Run the identical underlying setup with and without the candidate. Report lost winners, avoided losses, trade count, turnover, tail outcomes, and uncertainty.

Sizing task

Compare constant or naive risk sizing

Judge calibration, target-risk error, drawdown tails and concentration. A smaller return caused by lower exposure is not automatically worse.

Forecast task

Use a persistence or unconditional forecast

Score the prediction itself and any downstream trading application separately. A statistically better forecast may still lack executable value.

No random shuffling

Move Forward Through Time With a Frozen Family

6M observations are ordered, adjacent labels can overlap, and market states persist. Randomly mixing rows can leak later regimes and nearly identical paths into both training and test data. Walk-forward evaluation should reproduce what would have been known at each decision.

Develop

Choose the task, data audit, candidate family, parameter grid, baseline, costs, metrics and exclusions on early history.

Validate once

Apply the family to later dates. Document one controlled selection and every discarded candidate.

Lock the winner

Seal code, input hashes, parameters, roll policy, time rules, event set, costs and acceptance thresholds.

Open the holdout

Evaluate untouched later data once. A change after viewing it invalidates that holdout for the revised candidate.

Use purge and embargo intervals that cover the longest feature and outcome overlap. Re-estimation, if allowed, occurs on a fixed schedule using only the window available then. The full 6M backtesting protocol owns data lineage and simulation details.

Two bills arrive together

Pay for Trading Friction and Researcher Choice

Threshold-crossing rules tend to act when price or liquidity changes. A close-price fill ignores spread, market impact, latency and unfilled passive orders. At the same time, trying many correlated versions raises the evidence burden even when only one trades.

SpreadUse bid and ask appropriate to the order sideSlippageMeasure arrival-to-fill and stress tailsFees & rollsCharge every entry, exit, rebalance and expiry migrationMultiplicityDisclose every tested variant and control the family

Question	Required answer	Reject when
Could the order fill?	Order type, arrival quote, size, latency, depth or defensible queue model	A bar merely touched the price
What was tried?	Complete registry of formulas, parameters, filters, tasks and manual exclusions	Only the winner remains in the record
How was selection controlled?	Predeclared family-aware inference plus untouched chronological evidence	Nominal single-test threshold is applied to hundreds of trials
Is there cost headroom?	Performance under observed assumptions and materially worse stress	Point estimate disappears under a one-step cost change

Averages conceal dependence

Require Stability Across Regimes and Nearby Definitions

Do not demand identical performance everywhere; demand that the surviving claim matches the evidence. If value appears only in a named condition, that condition must be observable in real time, sufficiently sampled, stable later, and part of the rule.

Inputs

Change the construction

Adjacent bar boundaries
Second ex-ante roll rule
Nearby lookbacks
Alternate missing-data policy

States

Split without rescuing

Event/non-event
Calm/stressed volatility
Liquid/thin book
Early/later years

Outcomes

Inspect the full path

Effect size and interval
Drawdown and tail loss
Turnover and capacity
Missed-fill sensitivity

Scheduled controls should be built from official records such as the Banco de México publication calendar, INEGI release calendar, and Federal Reserve FOMC calendars. Their timing supports event labels; it does not establish a directional 6M response.

Ending decision

Indicator Decision Card

The card permits “none beat the baseline.” That is a completed research outcome, not a failure to choose a favorite.

Candidate: Task:

Definition: Formula, inputs, lookback, threshold, bar clock, signal time, expiration and warm-up are frozen.
Comparison: Task-matched naive baseline, identical risk, costs and eligibility, full family disclosed.
Validation: Chronological splits, overlap controls, untouched final holdout and access log are complete.
Execution: Dated-contract orders, spread, slippage, fees, rolls, latency, misses and partials are modeled.
Robustness: Effect survives reasonable adjacent definitions, later dates, regimes and cost stress.
Claim: Scope is no broader than observed task, data, size, horizon and conditions.

Accept for monitored useEvery required field passes and prospective monitoring has explicit retirement limits.

Keep experimentalEvidence is incomplete or fragile but the candidate remains worth a new, sealed test.

Choose noneNo candidate delivers stable incremental value after selection and implementation costs.

Research status as of August 13, 2026

No original study is reported. No original 6M indicator comparison was run for this article. The page names candidate families only to define a test. It does not recommend RSI, moving averages, ATR, volume, order-flow variables, or any combination as profitable or superior. Verify all current contract inputs in the canonical 6M contract-specification guide.

Sources, methods and editorial disclosure — reviewed August 13, 2026

CME Group Mexican Peso futures contract page, 2026 FX Product Guide, and CME Rulebook Chapter 256 for current exchange-product and dated-contract context.
CME DataMine historical-data catalog and CME Market by Order FAQ for trade, quote, depth and order-level data distinctions. No dataset was purchased or downloaded for this page.
NIST/SEMATECH guidance on multiple comparisons for the need to account for a family of comparisons.
NIST/SEMATECH introduction to time-series analysis and model-validation guidance for temporal dependence and diagnostics beyond one fit statistic.
CFTC advisory on commodity trading systems and hypothetical performance for hindsight and simulation limitations.
Banco de México 2026 publication calendar, INEGI 2026 release calendar, and Federal Reserve FOMC calendars for event-control sources.

Sources and methods were reviewed August 13, 2026. This page is unsponsored editorial analysis and presents a benchmark-and-rejection protocol, not original indicator findings.