Empirical impact of averaging/down-sampling order versus IID resampling in performance estimation

Determine the empirical impact on out-of-sample performance estimation of choosing between two pipelines when evaluating rolling-window portfolio rules on temporally dependent financial returns: (i) first average (smooth) or down-sample (decimate) the data and then apply IID resampling, versus (ii) first apply IID resampling and then average or down-sample. Quantify how these alternative orderings affect bias and variance of performance estimators such as the Sharpe Ratio.

Background

The paper studies bias introduced by IID resampling in backtests of rolling-window mean-variance portfolios, emphasizing that disrupting temporal dependence can create systematic errors. Beyond pure resampling, practitioners often smooth or down-sample financial data, and the order in which smoothing/decimation and resampling are performed can plausibly alter dependence structures differently.

The authors explicitly note that the empirical effects of these orderings are not established, motivating a targeted investigation into whether one pipeline provides performance estimation benefits over another and under what conditions.