Time-Series Models Overstate Sample Size Due to Overlapping Windows

A technical analysis highlights a subtle but critical flaw in time-series model validation: overlapping rolling windows create near-duplicate data rows that inflate the apparent sample size. For example, a 3,300-row dataset with a 30-day forward label yields roughly 110 truly independent observations, not the 3,275 rows the model sees during training. This inflation causes confidence intervals to be far too narrow, statistical significance tests to produce false positives, and hyperparameter selection to reflect random luck rather than genuine performance. Standard leakage checks and time-based train-test splits do not detect this issue, since every row remains causally valid and no future data crosses the split boundary. The author argues that practitioners must calculate and report effective sample size — row count divided by window length — to produce meaningful uncertainty estimates in temporal modelling.