Calendar hypotheses · multiple testing · sealed holdout
6N Seasonal Patterns: Test Calendar Effects Without Forecasting
Twelve months, four quarters, five weekdays, several month-end definitions and many holding windows create hundreds of possible 6N comparisons. Even random returns will make some cells look special. The first seasonal control is therefore not a chart; it is a record of how many questions were asked.
- Primary hypotheses
- Predeclared
- Roll policy
- Frozen
- Validation
- Chronological
- Original findings
- None
Freeze the question
Build a Hypothesis Registry Before Viewing the Calendar Table
A seasonal claim must identify its economic unit, comparison, direction, return window, entry and exit timestamps, eligible contracts, controls, outcome metric and rejection rule. Narrative intuition may motivate the test but cannot define it after results appear.
Example protocol—not a result
Test whether a single preregistered calendar window has a different distribution of costed 6N returns from matched non-window observations, using dated contracts, a frozen roll rule, block-aware uncertainty and a later untouched holdout.
- Fact
- Calendar labels are known in advance
- Hypothesis
- One label may condition returns
- Valid conclusion
- No effect
| Registry field | Example form | Prohibited ambiguity |
|---|---|---|
| Calendar unit | Named month, weekday, business-day index or event-relative window | “Around year-end” |
| Return | Exact completed-bar endpoints and direction | Choose close or intraday after inspection |
| Comparison | Matched ordinary observations or unconditional distribution | No baseline |
| Family | All months/windows/filters tested in the project | Publish only the winner |
| Decision | Effect-size and uncertainty threshold plus cost stress | “Statistically interesting” |
Count cycles, not rows
Define the Sample and Its Independent Seasonal Repetitions
Five-minute observations inside one January do not create thousands of independent Januaries. Report the number of years and eligible cycles, regime coverage, missing intervals and structural breaks alongside the number of raw bars.
A monthly hypothesis repeats once per eligible year, not once per intraday bar.
Disclose holidays, data gaps, delivery constraints and incomplete windows.
Rates, volatility and market structure can differ across eras.
Retain source files, revisions, calendars, transformations and checksums.
RBNZ’s B1 series provides official daily NZD exchange-rate and Trade Weighted Index data. It can support contextual comparison, but it is not a substitute for executable 6N futures returns or a continuous-contract policy.
Dated contracts first
Construct Returns Without Turning Rolls Into Seasons
6N is a deliverable, dated futures contract under CME rules. A recurring quarterly roll can line up with calendar labels and manufacture an apparent seasonal pattern if the series splice or transaction cost is mishandled.
Retain expiries
Ingest each dated contract and keep its prices, quotes, volume and delivery month separate.
Choose ex ante
Select the active contract using only information available before the migration decision.
Charge the roll
Model closing and reopening at executable prices with spread, slippage and fees.
Label afterward
Attach calendar bins only after return and roll construction is locked.
costed return in registered windowvs.costed return in predeclared comparisonBack-adjustment can remove a visual gap by changing historical levels. Preserve the mapping from every analytical return to the dated contracts and prices an order could have used.
CME Rulebook Chapter 258 defines the contract and delivery framework. Full mechanics remain in 6N Contract Specifications.
Calendar is not cause
Control Rival Explanations Without Inventing a Story
A repeated month can coincide with monetary-policy schedules, statistical releases, holidays, thin books, export reporting or broad-dollar regimes. These are candidate mechanisms to test, not proof that the month itself causes returns.
Scheduled information
Tag exact RBNZ, Stats NZ and U.S. release times. Test event versus non-event observations instead of assigning all movement to a month.
Market calendar
Version holidays, early closes, pauses and daylight-saving transitions. A short session changes opportunity and liquidity.
Contract lifecycle
Separate roll-proximate and ordinary periods by dated contract. Check whether the effect survives an alternate ex-ante roll rule.
Macro regime
Use only predeclared, observable conditioning states. Do not choose a regime split because it restores the desired result.
Stats NZ imports-and-exports documentation distinguishes monthly merchandise trade and quarterly overseas trade indexes. Its overseas merchandise trade metadata also documents provisional revisions. Release-time versions, not final revised values, belong in a causal historical test.
Small samples need honest intervals
Report Distributions, Tails and Family-Wide Uncertainty
An average month return can be dominated by one crisis year. Publish the median, dispersion, worst observations, sign count, full distribution and an uncertainty method that respects time dependence. Show the complete candidate family.
Magnitude
Effect size
Difference in preregistered location or distribution metric, before and after conservative costs.
Precision
Interval
Block-aware interval or resampling distribution with block choice and assumptions disclosed.
Tail
Adverse path
Worst cases, drawdown path, gaps and concentration by year or event state.
Search
Multiplicity
Total hypotheses, transformations and filters tried, with a family-aware correction or false-discovery control.
- Do not equate sign count with certainty. Years can have unequal magnitudes and common shocks.
- Do not choose the resampling rule after seeing significance. Freeze it with the hypothesis.
- Do not round away instability. Publish estimates at all registered robustness settings.
- Do not hide economic weakness. A narrow positive estimate that loses to costs is not a trading application.
Discovery is not confirmation
Freeze the Candidate and Open a Later Holdout Once
Calendar patterns are especially vulnerable to selection because the labels are easy to scan. Use early data for one documented discovery and later chronological data for confirmation. Any post-holdout change creates a new hypothesis.
Register
Claim, direction, timing, family, controls, costs, metrics and rejection rule.
Develop
Fit cleaning and limited choices in early data; retain all variants and failures.
Lock
Seal code, inputs, transformations, thresholds and access log.
Confirm
Open later data once and publish pass, limited or reject without rescue.
The CFTC advisory on commodity trading systems explains that hypothetical results have inherent limitations and may not reflect actual fills or a trader’s capacity to bear losses. Even a confirmed seasonal backtest remains simulated evidence.
From description to decision
A Calendar Effect Is Not Yet a Trading Rule
To become a trading application, a surviving effect needs a decision time, order type, size, liquidity gate, invalidation, exit, roll handling and risk budget. Compare it with a risk-matched naive rule and with no trade.
Fact
The calendar label and official event schedule were known by the declared decision timestamp.
Empirical observation
A separately run study estimates a distribution difference, with uncertainty and all searched variants disclosed.
Mechanism hypothesis
A declared event, flow or liquidity channel could generate the effect and survives targeted rival controls.
Trading application
A frozen, later-tested order rule adds decision value after spread, slippage, fees, misses and risk constraints.
Falsification decision
Seasonality Falsification Matrix
Record every row. A claim that survives only one definition or one exceptional year is rejected, not rebranded as a tendency.
| Test | Pass condition fixed in advance | Reject or limit when |
|---|---|---|
| Primary definition | Registered effect and interval meet the economic threshold | Wrong sign, too small or too uncertain |
| Alternate roll | Direction and useful magnitude survive another ex-ante policy | Effect is concentrated in splice or roll dates |
| Leave-one-year-out | No single year creates the conclusion | Removing one cycle reverses or erases it |
| Rival controls | Effect remains after registered event and regime controls | Calendar label proxies for another observable state |
| Multiplicity | Survives the declared family-wide procedure | Only unadjusted winner looks persuasive |
| Holdout and costs | Later data and stressed implementation retain decision value | Fails out of sample or after costs |
No original study or result is reported here: no 6N calendar sample, monthly tendency, weekday effect, seasonal coefficient, forecast, backtest or trading result. The page supplies a test protocol only. The current conclusion is not tested.
Sources, methods and editorial disclosure — reviewed August 20, 2026
- CME Rulebook Chapter 258: New Zealand Dollar/U.S. Dollar futures for dated-contract and delivery context.
- CME DataMine historical-data catalog for available exchange-data categories. No market dataset was purchased or analyzed for this article.
- Reserve Bank of New Zealand B1 exchange rates and Trade Weighted Index for official NZD reference-series definitions.
- Stats NZ imports-and-exports topic documentation and overseas merchandise trade metadata for release frequency, concepts and revision controls.
- U.S. Census Bureau X-13ARIMA-SEATS seasonal-adjustment documentation for official seasonal-model context; this article does not apply X-13 to 6N.
- CFTC advisory on hypothetical commodity-trading results for simulation limitations.
Sources and methods were reviewed August 20, 2026. This unsponsored article separates calendar facts, candidate mechanisms, registered hypotheses, inferences that require an actual study and trading applications that require later execution evidence.