Frozen rule registry · persistence baseline · walk-forward rejection
6S Mean Reversion: A Costed, Walk-Forward Test
The same 6S move can be called an overextension by a ten-bar trader and a trend by a hundred-bar trader. Neither label exists in the price alone. The anchor, lookback, threshold, entry delay and exit horizon create the claim. Mean reversion becomes testable only when those choices are frozen before the later data are opened.
- Candidate
- Fully specified
- Rival
- Persistence + neutral
- Costs
- Order-level
- Original test
- None reported
No discretionary rescue clauses
Write the Entire Candidate Rule Before Testing
“Buy an oversold 6S” is not reproducible. A registry must determine what data are eligible, when the state exists, when an order can first arrive and how every path ends. Parameter values are research choices, not recommendations.
| Field | Required declaration | Leakage to prevent |
|---|---|---|
| Instrument | Dated 6S expiry, roll policy, price field and session | Choosing the cleaner continuous series after results |
| Anchor | Past-only mean, median, VWAP or model state with exact lookback | Moving the anchor because price kept trending |
| Extension | Distance and scale estimator computed only from available history | Using full-sample volatility |
| Entry | Eligibility, trigger, confirmation, earliest order time, side and order type | Filling on the bar that created the signal |
| Risk | Size, maximum loss, price/time invalidation and event policy | Widening the stop after entry |
| Exit | Anchor target, partial rule, time stop, final exit and re-entry policy | Calling any profitable endpoint the target |
Eligible
Contract, event, liquidity and data gates pass.
Extended
Frozen displacement rule is true on completed data.
Triggered
Registered confirmation, if any, occurs without hindsight.
Ordered
Submission follows the declared delay and order policy.
Resolved
Fill, miss, exit, cost and invalidation are recorded.
Stabilization must be a timestamped, past-only condition. If two independent implementations can disagree about whether it occurred, the candidate is not ready for a backtest.
Reversion must beat plausible rivals
Benchmark Against Persistence, No Trade and Neutral Timing
A positive gross average does not prove reversion. The same eligibility filter may select favorable periods, or the rule may simply hold a standing CHF exposure. Compare risk-matched alternatives under the same timing and cost conventions.
Baseline 1
No trade
Zero exposure and zero trading cost. Complexity must justify taking risk at all.
Baseline 2
Persistence
Use the same extension event but trade continuation with symmetric risk and execution.
Baseline 3
Direction-neutral
Preserve eligible timestamps and holding periods while balancing or randomizing side without future data.
Baseline 4
Simple exposure
Compare with a risk-matched static or session-only CHF/USD exposure where relevant.
Primary question—not a finding
Does the frozen reversion candidate add net, risk-adjusted decision value over these baselines in chronological data never used to select its anchor, threshold or exit?
- Signal family
- Mean reversion
- Rival family
- Persistence
- Allowed winner
- Neither
The path to the anchor must be executable
Charge the Candidate for Every Order Outcome
Mean-reversion rules often respond to fast dislocations, precisely when spread, depth and adverse selection may be poor. Model marketable and resting orders separately. A missed limit order is an outcome, not a free fill at the desired price.
| Cost layer | Required treatment | Failure mode |
|---|---|---|
| Spread | Historical bid/ask at plausible order arrival or conservative spread model | Midpoint fills during stressed extensions |
| Impact | Size-aware book sweep or conservative impact schedule | Unlimited capacity at the inside quote |
| Latency | Decision, processing, routing and exchange acknowledgement delay | Entry before the signal was knowable |
| Limits | Queue-aware fills where possible; otherwise pessimistic bounds | Every touched price counted as filled |
| Misses | Opportunity cost, later chase rule or no-trade resolution | Keeping only successful fills |
| Fees and roll | Broker, exchange, clearing, market-data and roll costs as applicable | Gross performance labeled net |
net USD = realized USD P&L − spread USD cost − impact USD cost − fees USD − roll USD cost|misses and partials remain in sampleCME DataMine distinguishes settlement, Market by Order and PCAP datasets. That distinction matters because a bar-based test cannot claim queue-aware execution. No CME market data was purchased or analyzed for this page.
Temporary pressure can become repricing
Predeclare Regimes That Can Break the Anchor
A reversion story is most vulnerable when the information set changes and the old center is no longer relevant. Regime labels must be observable at entry or used only for ex-post diagnosis—never as hindsight filters.
Scheduled policy state
FOMC or SNB decisions, projections and official communication can reprice the relative policy path.
Unscheduled shock
Risk, funding or intervention-related information can move CHF while the historical anchor lags.
Liquidity state
Holiday, handover, benchmark and event windows can change spread, depth and fill behavior.
Contract lifecycle
Roll migration and expiry proximity can distort a continuous series or change execution capacity.
Volatility state
Past-only scale estimates can jump after the signal and invalidate a threshold calibrated to quiet data.
Dollar-factor state
A broad USD move can create persistent cross-currency pressure unrelated to a 6S-specific anchor.
Optimization ends before validation begins
Use Nested, Chronological Walk-Forward Testing
Time order is part of the strategy. All parameter choice, regime thresholds and model fitting happen inside training history. The next block is evaluated once, then appended only after its result is sealed.
Register candidates
Limit anchor, lookback, threshold, confirmation, exit and regime families before outcome review.
Select inside history
Use nested resampling or an inner chronological split for every parameter decision.
Open the next block once
Apply frozen rules with causal data, latency, costs, partials and misses.
Advance the origin
Log parameter stability and every block result; never overwrite a failed generation.
Final holdout
Use once for the surviving registered candidate and its baselines.
- Family-aware uncertainty. Count every tried anchor, threshold, horizon and regime rule.
- Block dependence. Use uncertainty methods consistent with overlapping trades and clustered events.
- Capacity stress. Re-run at larger spreads, delayed entry and lower fill probability.
- Subperiod disclosure. Show whether one episode carries the result.
Failure is a valid product outcome
Abandon the Mean-Reversion Claim When Evidence Fails
A candidate does not earn indefinite retuning. Set rejection conditions before the final holdout and apply them without a discretionary rescue.
Statistical
Unstable edge
- Sign or useful magnitude changes across later blocks
- Interval includes economically harmful outcomes
- One event or regime dominates the estimate
Economic
No net value
- Costs consume gross reversion
- No-trade or persistence baseline matches it
- Capacity is below practical order size
Operational
Cannot reproduce
- Required quotes, events or rolls are missing
- Live state cannot be identified without hindsight
- Fill assumptions cannot be defended
Final strategy record
- Definition
- Rule version, data manifest, code hash, sample cutoffs and cost schedule.
- Search
- All candidates, rejected variants and selection process.
- Evidence
- Block-by-block gross and net outcomes, uncertainty, drawdowns and capacity.
- Accept
- Stable incremental value survives the sealed holdout and conservative execution.
- Limit
- Use only in a real-time-identifiable state with documented size constraints.
- Abandon
- Any preregistered statistical, economic or operational failure is triggered.
No original 6S mean-reversion backtest, trade count, hit rate, Sharpe ratio, drawdown, parameter or net performance is reported here. The claim that 6S “is built for mean reversion” is not established. This page defines how a candidate can be tested and rejected.
Sources, methods and editorial disclosure — reviewed August 21, 2026
- CME Rulebook Chapter 254: Swiss Franc/U.S. Dollar futures for dated-contract identity and exchange rules.
- CME Swiss Franc futures contract page for current product context; specification arithmetic belongs to the canonical 6S contract guide.
- CME DataMine historical-data catalog for distinctions among settlements, Market by Order and PCAP datasets.
- NIST/SEMATECH time-series analysis guidance for time order, dependence and model-checking context.
- Federal Reserve FOMC calendars and official materials for U.S. policy-event controls.
- Swiss National Bank data portal for official Swiss rate, yield and foreign-exchange context.
Sources and methods were reviewed August 21, 2026. This unsponsored article separates contract facts, proposed mechanisms, candidate rules, statistical evidence and trading use. It reports no original market result.