Definitions · decision clock · transitions

ES Market Structure: Forward-Safe Trends, Pullbacks, and Reversals

By Kyle ParrottUpdated August 25, 2026

Show the same ES path to two analysts and one may label a healthy pullback while the other sees the start of a reversal. The disagreement is not solved by drawing more lines. It is solved by defining observable states, the time each label becomes available and the evidence that retires it.

Define the states Permit unusable

Features: Past and present only
Label time: Explicit
Transitions: Estimated, not promised
Hindsight: Blocked

States need necessary conditions

Define Trend, Pullback and Reversal as Observable Rules

Choose a price series, bar interval, session, contract-roll policy and lookback. Compute all features from completed observations. Thresholds should be scaled by returns, true range or another preregistered denominator so the same rule can be evaluated across price levels.

State	Necessary context	Candidate observable rule	What it does not prove
Up state	Declared trailing window	Positive normalized slope or net displacement plus minimum directional efficiency	That the next return is positive
Down state	Same feature family and horizon	Negative normalized slope or displacement plus the same efficiency gate	That selling is profitable after costs
Pullback candidate	Pre-existing directional state	Counter-state retracement inside declared depth/duration bounds without invalidation	That the prior state will resume
Reversal candidate	Pre-existing state or extreme	Inactivation of old state plus opposite-state evidence available now	That later follow-through is already known
Neutral/uncertain	Any period	No directional rule passes, rules conflict or data fail	That the market is low risk

One path-efficiency featureE_t,h = |P_t − P_t-h| / Σ_i=t-h+1^t|P_i − P_i-1|E approaches one for a direct path and zero for heavy back-and-forth. It describes the completed window; it is not a forecast.

A chart label has an availability timestamp

Separate Detection, Confirmation and Outcome

A common leak occurs when a swing high is defined only after several lower bars, then treated as known at the high. Preserve three clocks so research cannot move information backward.

Observation

The latest eligible quote, trade or completed bar arrives at source time t.

Feature

Inputs through t are computed under the frozen transform and session rules.

Candidate

A state becomes actionable only after the system can calculate it.

Confirmation

Later observations may confirm or reject the earlier candidate, never rewrite its decision time.

Outcome

Forward returns and fills are scored after latency and costs.

Zigzags and centered extrema are labels, not live signals.

If a pivot requires future bars, it can be used as a retrospective outcome for supervised research. It cannot be supplied to a decision at the pivot timestamp. Store both pivot time and label-availability time.

Measure what followed each frozen state

Build a Transition Matrix With Honest Uncertainty

For each eligible time, assign one state using only then-available data. At a fixed later horizon, record the next state. Estimate conditional transition distributions by session, volatility, event and roll regime only when sample sizes support them.

State at t	Up at t+h	Neutral at t+h	Down at t+h	Required report
Up	Continuation count/rate	Decay count/rate	Opposite-state count/rate	n, missing, uncertainty
Neutral	Up-transition count/rate	Persistence count/rate	Down-transition count/rate	n, missing, uncertainty
Down	Opposite-state count/rate	Decay count/rate	Continuation count/rate	n, missing, uncertainty

The table above is a reporting template, not an ES finding. Use non-overlapping decision points or account for overlapping outcomes. Keep state dwell time, maximum excursion and executable returns separate from state-transition rates; they answer different questions.

Pullback versus reversal as a sequence

Suppose an up state exists at t₀. A counter-move at t₁ passes the pullback-candidate bounds. At t₂, either the prior up-state condition reactivates, the candidate remains unresolved, or the old state invalidates and an opposite state qualifies. The t₂ result can score the t₁ candidate. It cannot be used to claim the trader knew “pullback” or “reversal” at t₁.

A label must be stable enough to use

Validate Definitions Before Testing a Trade

First test whether humans and code can apply the definitions consistently. Then test state stability across reasonable sampling choices. Only after those gates pass should a separate execution strategy be evaluated.

Reproducibility

Independent relabel

Two implementations receive the same frozen data and specification. Report agreement and inspect every conflict.

Sensitivity

Reasonable perturbations

Vary bar boundaries, threshold edges and feed source within preregistered ranges. A useful label should not flip everywhere.

Forward test

Untouched dates

Lock definitions after training. Score transition calibration and any execution rule on later observations once.

Economics

Executable costs

Use order side, latency, spread, depth, slippage, fees and rejected signals. A correct label need not produce an edge.

CME provides official ES contract context and historical market-data categories through its ES product page and data catalog. The state definitions on this page are proposed research constructs, not exchange labels.

A failed labeling system is a valid result

Retire the Model When Structure Is Not Operationally Clear

Do not force every path into trend, pullback or reversal. The research is unusable if labels depend materially on hindsight, independent implementations disagree, thresholds are unstable, live states fall outside training support or economic results fail after costs.

Semantic failure

Definitions still require discretionary phrases such as “clean” or “obvious.”

Action: rewrite or stop

Timing failure

Label availability occurs after the assumed decision price.

Action: invalidate the test

Economic failure

Forward transitions do not survive executable costs and uncertainty.

Action: no trading claim

Every feature has a source timestamp and availability timestamp.
State definitions include neutral and data-failure outcomes.
Threshold selection is confined to the training sample.
Transition tables report counts, missingness and uncertainty.
Event, session, volatility and roll regimes are prespecified.
Retrospective labels never masquerade as live decisions.

Research status as of August 25, 2026

No original ES trend persistence, pullback success rate, reversal rate or after-cost strategy result is reported. This article supplies a forward-safe labeling and validation protocol only.

Sources, methods and editorial disclosure — reviewed August 28, 2026

CME E-mini S&P 500 product and contract page for ES identity and exchange context.
CME futures and options data catalog for distinctions among trade, top-of-book, depth and market-by-order data.
CME Equity Quarterly Roll Analyzer guide for treating contract migration as a separately observable state.
S&P U.S. Indices Methodology for the underlying index framework.
CME trading-hours and holiday schedule for session-integrity controls.

Sources were reviewed August 28, 2026. All market-structure states and formulas are disclosed research choices, not exchange classifications or completed findings.