[throwaway161803]

Intraday data is often unnecessary and, worse, leads otherwise intelligent people to trade on spurious relationships. Recall that a sufficient statistic for the drift of a Brownian (or geometric Brownian) price process over a period is its return over the period. Hence intraday data provides essentially no additional information on the drift of prices, though it is useful for measuring the strength of covariate relationships and responses to intraday events.

[lukebuehler]

I did not say that the overall model must be built on intraday data.

But most models will simply be useless without proper execution. And IMO, the model should be back-tested with simulating the transaction costs as close as possible.

So the overall model might be fine and interesting, but if you can't close the loop with a proper execution model, which usually affects the overall model as well, you simply don't have much.