In this dissertation, I propose a new model for the analysis of financial durations. The new model improves upon several limitations of the autoregressive conditional duration (ACD) model considered in Engle and Russell (Econometrica 66(5) (1998) 1127-1162). Instead of adopting the multiplicative error form assumed by the ACD model, I establish a mixture of exponentials representation for durations from general point process theory. Based on the representation, I develop the Markov switching multifractal duration (MSMD) model. I present the geometric ergodicity property of MSMD and show that the MSMD can explain most stylized facts of financial durations, especially the long memory feature. An extensive empirical study shows MSMD compares favorably with ACD both in- and out-of-sample. For long horizon forecasting, MSMD dominates ACD, which confirms that MSMD can explain long range dependence in durations.