This paper analyzes the performance of different models and extracts periodic patterns from the transactions in TARGET2, the Eurosystem's real-time gross settlement system. We present a "horse race"-style comparison of (1) a classic autoregressive- moving-average (ARMA) model with dummies, (2) an ARMA with trigonometric seasonal cycles and (3) a TBATS state space model introduced by De Livera et al. The models investigate different layers of the network (interbank, customer and money market payments) and the individual bank transactions of four Dutch banks. The periodic cycles included in our models range from intraday to intrayear. Our results show that the level of granularity coinciding with the best model fit depends on the transaction type: (1) ten-minute aggregates for customer and money market transactions, and (2) one-hour aggregates for individual banks and interbank transactions. The performance of forecasts exhibits much greater variation between models. However, the more granular (ten-minute) level outperforms the less granular (one-hour) level differences in most cases. Further, we find that the classic ARMA model with dummies extracts significant cyclical patterns from individual bank transactions, which is a useful result for supervisors. However, the TBATS model, which allows for cyclical components that vary across time, allows for a more precise analysis. Finally, the cyclical patterns of individual banks (may) differ greatly.