Operational definition · event labels · validation
6B Pullbacks: Define and Test a Clean Retracement
A testable pullback is a timed sequence: a causally identified impulse, a bounded retracement measured from that impulse, a completion rule known at the decision, and a separately defined forward outcome. Chart scale, bar interval, overlap tolerance and execution costs all belong in the frozen record.
Measurement designRules before charts
The contradiction
A Clean Pullback Begins With Frozen Impulse and Retracement Rules
Fact: 6B is a standardized CME British Pound futures contract quoted in U.S. dollars per pound. Observation: traders often call a retracement “clean” when price gives back part of a prior move without obvious overlap or violent reversal. Hypothesis: some predeclared versions of that shape may condition later price paths. The hypothesis still needs a dataset and an out-of-sample test.
A usable event definition
Start an impulse only when a past-data rule is met—for example, a signed move of at least k times the trailing median true range within no more than m bars. End it at the first qualifying swing extreme produced by a frozen swing algorithm. A pullback then begins after that extreme and ends only when a separately declared termination rule fires.
- Instrument
- Dated 6B contract
- Clock
- Exchange timestamps in UTC
- Decision point
- Known in real time
- Claim status
- Unvalidated hypothesis
A swing label that needs bars to the right can be used for retrospective description, but not as a live entry time unless the confirmation delay is included. Likewise, a ZigZag that redraws after new extremes introduces look-ahead if its earlier turning point is treated as contemporaneously known.
The standard contract represents 62,500 British pounds and the outright minimum increment is 0.0001, worth $6.25 per contract. A five-tick move is $31.25 before fees and slippage. See the canonical 6B contract guide for mechanics; those mechanics do not validate any pullback rule.
Make the drawing measurable
Turn “Clean” Into Observable Variables
Do not hide several judgments inside one yes-or-no label. Store the components first. A later model can test whether depth, path, time, and market quality matter independently.
| Variable | Reproducible definition | What it describes | Main bias risk |
|---|---|---|---|
| Impulse size | Signed ticks from frozen start to frozen extreme | Scale of the move being retraced | Choosing the start after seeing the endpoint |
| Retracement depth | Absolute pullback ticks divided by absolute impulse ticks | Fraction of the impulse surrendered | Changing the acceptable band by example |
| Path efficiency | Net pullback distance divided by the sum of absolute bar-to-bar changes | How direct or overlapping the route was | Bar-size sensitivity and bid-ask bounce |
| Duration | Elapsed seconds and number of fixed bars | Time spent retracing | Mixing active and quiet sessions |
| Adverse extension | Maximum distance beyond the declared invalidation before the horizon ends | Tail risk after a candidate appears | Using a future extreme to move the stop |
| Participation | Executed 6B volume relative to the same clock window | Exchange-traded activity, not total OTC GBP flow | Calling volume an actor or motive |
| Market quality | Quoted spread, displayed depth, trade-throughs, and missing-message flags | Whether the observed path was executable | Using completed bars without quote data |
| Context labels | Session, official release, holiday, contract, and days to roll | Known differences in information and liquidity | Adding labels only after a surprising result |
Illustrative calculation, not a finding: if a declared 30-tick impulse is followed by a 12-tick net retracement, depth is 40%. If the one-minute path travels 18 ticks in total to produce that 12-tick net move, path efficiency is 0.67. Different bar lengths can change the second number, so report a robustness grid.
Event construction
Build the Sample Without Looking Past the Decision
The observation unit should be one eligible impulse-pullback event, not one handpicked chart. Preserve every qualifying event, including overlapping signals, failed retracements, trading halts, thin periods, and occasions when the rule could not obtain a fill.
Freeze data
Use dated contracts, a documented roll rule, UTC timestamps, and an auditable bad-tick policy.
Detect impulse
Apply one past-only size and time rule to every eligible bar.
Label pullback
Record continuous variables before assigning any clean/not-clean class.
Lock outcome
Measure the forward path from the first moment the label was actually knowable.
Dependency control
Overlapping events are not independent
If several candidate pullbacks share the same impulse or forward horizon, naive standard errors overstate effective sample size. Keep a non-overlapping primary sample or use dependence-aware resampling by trading day or event block.
Contract control
A continuous chart can manufacture a move
Store the actual security and roll decision. Exclude or separately report stitch bars. A volume-based front-contract rule must use only volume available before the decision, not the final daily leader.
Split chronology before choosing thresholds: development, validation, and final holdout. If many depth bands, bar sizes, sessions, or termination rules are tried, preserve the entire search universe and adjust the inference for multiple testing. The attractive survivor is not the only experiment that happened.
Counterfactuals
Compare Against the Right Baselines
“Price resumed after the pullback” is incomplete. The same continuation may occur after any sufficiently large move, and a high win rate can coexist with poor tail risk.
Compare labeled pullbacks with every qualifying impulse, including those with messy or no retracement.
Match direction, size, time of day, volatility state, event status, and days to roll.
Test candidates just inside and outside each cutoff. A sharp result at one arbitrary boundary is fragile.
Apply the same measurement to nearby nonevent times where no candidate existed.
Compare with a rule based only on impulse size or momentum so “cleanliness” must add information.
Subtract commission, exchange and data fees where relevant, spread, slippage, and missed fills.
Report the full forward distribution: continuation frequency, median and tail maximum favorable excursion, maximum adverse excursion, time to outcome, and net results under more than one cost assumption. The 6B momentum workflow owns momentum-state decisions; this page asks only whether pullback variables add value beyond that state.
Trading application
Convert a Validated Label Into an Execution Question
Even a stable descriptive relationship does not specify an order. Entry timing, invalidation, size, and fill assumptions require their own predeclared rules.
A legitimate decision statement
“When the frozen pullback definition appears in a validated context, I will consider a limit or stop order only if the live spread and depth are inside the tested range, the invalidation is observable, and the cost-adjusted reward distribution still fits the risk budget.”
- Entry
- Exact trigger and order type
- Invalidation
- Price or time condition
- Friction
- Spread, queue, slippage
- Stand aside
- Event or liquidity gate
A limit order can miss the trade or fill mainly when the path continues against it. A stop order can fill beyond its trigger during a fast move. A completed-bar backtest that assumes the chosen price was always available ignores both problems. Evaluate marketable and passive variants separately, and never use the bar low as proof that a buy limit received a fill.
Failure conditions
Where the Pattern Fails or Disappears
A useful result must survive more than one attractive chart scale and more than one market regime.
Small, reasonable changes to swing, depth, or duration rules reverse the result.
Nearly all apparent benefit comes from one clock window or release category.
The gross relationship disappears under observed spreads, slippage, and missed fills.
Frequent small continuations are outweighed by rare, large adverse paths.
The rule works only in the development sample or after parameters are retuned.
The label changes materially between last trade, bid/ask midpoint, or data vendors.
If the result cannot survive the predefined holdout, neighboring parameters, event controls, and executable costs, retire the rule. Renaming a failed threshold is not validation.
Annotated research checklist
What Must Be Written Down Before the Next Chart
This is the page’s endpoint: a reproducible record that keeps a descriptive label separate from validated expectancy.
| Freeze before testing | Minimum record | Reason |
|---|---|---|
| Data lineage | Vendor, feed type, timestamp field, timezone, security IDs, and missing-data policy | Makes the path auditable |
| Impulse and pullback rules | Bar interval, thresholds, swing confirmation delay, termination, and overlap policy | Prevents hindsight drawing |
| Controls | Session, official releases, holidays, volatility state, contract roll, spread, and depth | Tests rival explanations |
| Outcome | Horizon, MFE, MAE, continuation, reversal, time, order model, and all costs | Shows the whole distribution |
| Validation | Chronological splits, search universe, robustness grid, and retirement criterion | Protects the holdout |
No original 6B pullback dataset, backtest, correlation, win rate or expectancy estimate is reported in this article. It supplies a measurement design. Any future result should publish sample dates, event count, exclusions, code or exact rules, controls, cost model, uncertainty intervals and holdout performance.
Sources, method and editorial disclosure
- CME Group British Pound product overview for contract size, quotation, tick, and trading-hour context.
- CME Group MDP 3.0 dissemination documentation for the distinction among bid, ask, trade, statistical, and recovery data.
- Andersen, Bollerslev, Diebold, and Labys, Modeling and Forecasting Realized Volatility for constructing realized measures from consistent intraday returns.
- CFTC basics of futures trading for leverage, contractual obligations, and loss-risk context.
Formulas and numbers are methodological or illustrative, not performance claims. Sources and time-sensitive contract facts were reviewed August 13, 2026. This is original, unsponsored editorial analysis.