The standard construction ranks on returns over the past twelve months excluding the most recent month, since very short horizons show reversal rather than continuation. The top decile is bought and the bottom decile sold, with the portfolio refreshed monthly.
Momentum has been found across equities, bonds, currencies and commodities, which makes pure data mining a less comfortable explanation. Its weakness is the crash: momentum portfolios tend to be short beaten-down, high-beta names, and when the market snaps back violently those shorts rally hardest. Losses of tens of percent in a month have occurred.
Turnover is also high, often several hundred percent a year, so implementation costs consume a large share of the paper premium. See trend-following for the managed-futures cousin of the same idea.
Related: factor-investing, trend-following, portfolio-turnover, factor-crowding, low-volatility-factor, mean-reversion