ECTO labels revealed a decision-time alignment problem

ECTO gave us a precise lesson about market labels: the price used as the return reference must match the decision being evaluated. Our historical target takes the logarithm of mean price in the following target window divided by the first price of the input window. That definition creates a valid research quantity, but part of the measured movement can occur during the input the model has already observed.

The DXRG aggregate audit discloses a 300-second input window and a 900-second target horizon. A model acting after the five-minute input faces a different return question from a label anchored at the input's opening. We learned to inspect that distinction before interpreting a class prediction as evidence about an entry and exit.

Decompose what the label contains

Call the input opening price P0, the price at decision time Pd and the following window's mean price M. ECTO's disclosed target is log(M/P0). The identity log(M/P0) = log(Pd/P0) + log(M/Pd) splits it into two components.

The first component measures movement from the input opening to decision time. It can already be visible in the input. The second measures the following window's mean price relative to the decision-time reference. These components can have opposite signs, so a correct prediction of the original label can answer a different question from the one a new position faces.

Consider illustrative prices: the input opens at 100, reaches 120 at decision time, and the following fifteen-minute window has a mean price of 115. The original label is log(115/100), approximately 0.1398. That exceeds ECTO's pump threshold of 0.12.

Yet relative to the decision price, log(115/120) is approximately -0.0426. An illustrative decision-time classifier using the same numerical bands would put that mean-price change in the flat range. The arithmetic explains how an original pump label can coexist with a lower future mean than the price available when the model decides.

This example identifies target mismatch rather than demonstrating a future-data leak. The problem is the reference time: observed movement contributes to the quantity scored as the target. Predicting persistence of the resulting class may still teach us about a token's state. A forecast of future movement needs its own reference and evaluation.

Class bands express a particular research question

ECTO's disclosed classes mark dump below a log return of -0.12, flat between -0.08 and 0.08, and pump above 0.12. Intermediate zones are ignored. These thresholds concern logarithmic returns, so 0.12 corresponds to approximately 12.75% in ordinary percentage terms, while -0.12 corresponds to approximately -11.31%.

The ignored zones matter because they shape the evaluated population. A classifier scored only on the retained bands answers a question conditioned on that selection. It needs an explicit treatment for intermediate outcomes if a later runtime must make decisions across the full stream.

Changing the reference to decision time can also change class membership and prevalence. The earlier flat-majority baseline cannot simply be copied into the repaired experiment. We would recompute labels, population counts and comparators together, preserving the original definition for historical comparisons.

This is why label repair is more substantial than changing a column name. It changes the problem the model is trained to solve and the distribution against which success is measured.

A future mean still differs from execution

Even log(M/Pd) remains a mean-price target. A trader usually receives a particular entry fill and exits according to a policy. Achieving the mean price of an entire future window would require a specified executable mechanism, including timing and liquidity. The mean itself supplies no fill receipt.

Our proposed successor starts with the decision timestamp, the data available by that timestamp and a clearly defined entry reference. It then chooses a future outcome suited to the question: a terminal price, a specified exit policy or a descriptive mean. Each is useful when its interpretation stays explicit.

A separate replay can ask whether the prediction supports executable actions. That replay needs entry delay, fees, slippage and available capacity, followed by account reconciliation. The historical reported target-label averages lack those execution checks, so they remain label statistics. Their sign or size cannot establish net trading returns.

The constructive outcome is a cleaner sequence of experiments. First ask whether the representation predicts future information beyond what the input already reveals. Then ask whether a frozen decision policy can use that prediction under realistic execution. ECTO's original target taught us exactly where those questions diverge, and gives us a concrete specification to improve before collecting stronger performance evidence.