Forecasting the pulse: how deviations from regular patterns in online data can identify offline phenomena