We replicate Chinco, Clark-Joseph, and Ye (2019) and investigate how sensitive return predictability results are with respect to how returns are constructed. We find that 1-minute returns are predictable only if returns are constructed using the last quote recorded in each one minute interval, irrespective of which exchange posted the quote, i.e., when returns capture both time-series and cross-sectional variation across exchanges. Predictability is largely due to spikes in prices, leading to strong negative autocorrelations. Return predictability is significantly lower if we remove spikes or when using other methodologies to construct returns based on prices without spikes.