Create stable identifiers and time keys
Every record needs a reliable entity ID plus a consistent time unit. Without that, joins and cohort comparisons become fragile.
Engineer change, not only levels
Useful features include rolling averages, slope, volatility, time since last event, threshold crossings and gaps in participation.
Aggregate at the decision level
The correct grain depends on the product question: student-week, customer-month, host-level or case-level features lead to different models.
Protect against leakage
Features used for prediction must only use information available at the decision time. Longitudinal work makes accidental future leakage especially easy.