Skip to main content

Orientation · multiple tests · costs · holdout

6Z Seasonal Patterns: A Holdout-First Calendar Test

A large average return for one month is easy to discover after trying twelve months, weekdays, month-end windows and election-year labels. It is also easy for one crisis year to create it. The only defensible seasonal claim begins with a frozen hypothesis, correct 6Z quote orientation, auditable contract rolls and a holdout that was never used to choose the pattern.

Direction
USD per ZAR
Tests tried
Count all
Costs
Mandatory
Claimed pattern
None

One claim at a time

Define the Calendar Effect Before Seeing It

State the entry and exit timestamps, timezone, return direction, eligible contracts, roll treatment and comparison set. “6Z is strong in spring” is not a hypothesis: spring differs by hemisphere, “strong” lacks quote orientation and no holding interval is specified.

Field Required declaration Example of ambiguity to remove
Calendar Month, business-day index or exact recurring window Calendar month versus final five trading days
Return Log or simple return on ZAR/USD 6Z quote Positive 6Z means a higher USD value per rand
Instrument Dated futures and explicit roll rule Vendor continuous series may back-adjust
Decision time Observable price and information timestamp Settlement not yet published at entry
Family Every related calendar test counted Reporting only the best of twelve months

Reconstruct what was tradable

Build a Point-in-Time Contract Sample

Use dated settlements or intraday prices and a roll policy based only on information available at the time. Preserve both raw contracts and the derived series. A back-adjusted chart can be useful for display but can alter price levels and returns around rolls; it cannot be accepted without its transformation being documented.

1

Acquire

Record source, fields, timezone, corrections and access date.

2

Validate

Check duplicate dates, missing sessions and anomalous settlement changes.

3

Roll

Apply the frozen volume, open-interest or calendar rule without future data.

4

Label

Attach official holidays and registered calendar windows.

5

Freeze

Hash the sample before exploratory results are viewed.

Worked hypothesis example

Turn “month-end strength” into a testable claim

Define month-end as the final five eligible exchange trading days and the comparison as all other eligible days in the same sample. Choose the entry and exit marks, ZAR/USD return sign, dated-contract roll rule and whether positions overlap before opening the data. The unit of inference should reflect dependence: daily observations inside one month-end window are not five independent calendar events. Aggregate by window or use an uncertainty method that preserves the clustering.

Report every yearly window, not just the pooled mean. Show median return, dispersion, tail outcomes, hit rate with uncertainty, turnover and conservative execution cost. Repeat the registered estimate after removing each year and after excluding separately declared SARB, FOMC, holiday and roll windows. Compare with a randomized calendar rule that preserves holding period and trade count. If the effect is carried by one crisis, changes sign under a defensible roll convention or fails the untouched period after costs, reject it. If retained, the claim remains limited to the exact window, quote orientation, sample and implementation tested.

Keep an experiment ledger with the registration timestamp, code revision, input hash and disposition of every related window. That record makes an abandoned hypothesis visible and prevents it from returning later as a supposedly new discovery.

Do not manufacture observations.

Missing sessions are missing. Do not forward-fill returns, treat absent quotes as zero movement or combine unlike serial and quarterly contracts without a documented reason.

The calendar may proxy for something else

Separate Calendar Labels from Regime and Event Exposure

A recurring window may coincide with SARB meetings, Federal Reserve decisions, holidays, tax dates, commodity cycles or contract migration. Predefine which are exclusions, which are stratification variables and which are rival explanations. Do not add controls only after they erase or strengthen the desired result.

Policy

Official event calendar

SARB and FOMC windows receive exact source timestamps and separate reporting.

Contract

Roll and expiry

Flag migration and exclude unsupported continuous-series seams.

Execution

Spread and capacity

Use time-matched cost evidence rather than a universal fixed deduction.

Make chance work harder

Correct for Search and Stress the Result

Publish the complete hypothesis family, not only its winner. Use a preregistered familywise or false-discovery procedure appropriate to the research question. Report effect size and uncertainty alongside any adjusted significance measure; a small p-value is not a trading edge.

Leave-one-year-out

Repeat the estimate after removing each year. A single-year effect is disclosed, not averaged away.

Neighboring windows

Shift entry and exit by registered increments. A one-timestamp spike is fragile.

Alternative rolls

Compare defensible roll policies without selecting whichever produces the best return.

Sign and rank tests

Check whether the result depends on the mean and extreme observations.

Cost stress

Apply spread, slippage, fees and missed-fill assumptions tied to the order policy.

Economic relevance

Compare net magnitude with risk, drawdown, capacity and a simple no-season baseline.

No second look

Lock the Rule Before Opening the Holdout

Choose the holdout boundary before model selection. Freeze the exact calendar rule, position direction, order convention, costs and pass threshold. Run once. If the rule fails, changing the dates or threshold creates a new research program and requires a new untouched sample.

train + choose freeze rule one holdout run retain or reject

Abandonment is a valid result

Reject a Pattern That Cannot Survive Its Costs and Controls

Retain a calendar effect only when its orientation, sample and search family are auditable; uncertainty excludes an economically trivial effect under the registered criterion; sensitivity checks do not reveal one-date dependence; and the untouched holdout clears conservative costs. Anything less remains exploratory.

Retain

Replicated after costs

The locked effect clears its holdout threshold with stable sign and usable capacity.

Monitor

Descriptive only

The calendar difference is measurable but too uncertain or small for a trading rule.

Reject

Fragile or unprofitable

It fails holdout, multiplicity, neighboring windows or conservative costs.

Unscorable

Sample is inadequate

Roll, missingness or available years cannot support the claim.

Research status as of August 22, 2026

No original result is reported. This protocol publishes no 6Z monthly, weekday, month-end or annual-cycle finding and does not assert that any calendar window predicts returns.

Sources, methods and editorial disclosure — reviewed August 25, 2026

Sources were reviewed August 25, 2026. This unsponsored article is a protocol, not a completed backtest, seasonality finding or recommendation.