A gold backtest can see tomorrow even with the rows in date order

Write down when the answer arrives

Suppose a hypothetical model makes a decision at the close of each 15-minute gold bar and tries to predict the return over the following hour. The label for a 09:45 UTC decision divides the 10:45 close by the 09:45 close and subtracts one. The inputs might all be available at 09:45. Its answer arrives an hour later.

Freeze the model immediately before 10:00 UTC, with the first test decision at 10:00. A rule that assigns every row dated before 10:00 to training includes the 09:45 label. The model would then have learned something about a price forty-five minutes beyond its freeze. This timeline uses hypothetical observation times and makes no claim about gold returns.

Move the boundary from the row to the evidence

Assume uninterrupted bars, instant availability at each close and the same one-hour horizon for every label. Under the strict freeze, 08:45 is the last eligible decision because its answer arrives at 09:45. The 09:00 label arrives at exactly 10:00, too late for a model already frozen. Decisions at 09:15, 09:30 and 09:45 must also wait. Those four rows can be used in a later training round once their outcomes are known.

A time-series splitter can leave a gap measured in samples. Four samples implement the boundary in this particular regular grid. Missing bars, mixed timeframes or a target defined by the first exit can change how much clock time four rows represent. Store an explicit outcome-availability time for each decision and check it against the actual freeze time.

Source notes: [1] scikit-learn — TimeSeriesSplit

Test the whole fit

Apply the same cutoff to everything learned from the sample, including a volatility scaling factor, an imputation value or a decision threshold. Fitting a scaler on the full export lets later observations change earlier inputs even if every forward label is correctly excluded. Fit the transformations on the eligible training slice and reuse those fitted values for the test decisions.

Source notes: [2] scikit-learn — common pitfalls and data leakage

Try a deliberate disturbance

I would make a copy of the dataset and replace the 10:45 close with an extreme artificial value. Rebuild only the training data and frozen model using fixed random seeds. Their inputs, eligible labels and parameters should be identical. The test-period outcome that legitimately uses the changed price can change. If the frozen model changes, trace the disturbance through labels, transformations and parameter selection before interpreting any score. This check finds a specific dependency. Passing it leaves other sources of leakage to investigate.