And plenty of studies show reversal at the week level, so a bad last week is good.

True enough.
But, anecdotally: though this may be true on average, I find it is so unreliable that I have removed the negative one-week factor from all my screens. There seem to be multi-year stretches that good one week performance is good, not bad, for the next month.

The better improvements to momentum I found were (again, anecdotally from trading and backtesting)--
* the optimum length of lookback seems to vary with a stock's turnover ratio. i.e., the optimum look back seems to be some constant K times the average trader's hold period.
* lookbacks work better for some classes of stock with a LOT of lag than they do with fresh data, especially long lookbacks. i.e., instead of a negative factor on the most recent week (or more), just ignore the most recent week (or more). Even up to six weeks of lag for some boring value/dividend stuff.

I don't know if those comments have any merit, but I thought I'd share 'em.

Jim