1: \begin{abstract}

2: 

3: We present a new algorithm for discovering patterns in time series and other

4: sequential data.  We exhibit a reliable procedure for building the minimal set

5: of hidden, Markovian states that is statistically capable of producing the

6: behavior exhibited in the data --- the underlying process's \emph{causal

7:   states}.  Unlike conventional methods for fitting hidden Markov models (HMMs)

8: to data, our algorithm makes no assumptions about the process's causal

9: architecture (the number of hidden states and their transition structure), but

10: rather infers it from the data.  It starts with assumptions of minimal

11: structure and introduces complexity only when the data demand it.  Moreover,

12: the causal states it infers have important predictive optimality properties

13: that conventional HMM states lack.  We introduce the algorithm, review the

14: theory behind it, prove its asymptotic reliability, use large deviation theory

15: to estimate its rate of convergence, and compare it to other algorithms which

16: also construct HMMs from data.  We also illustrate its behavior on an example

17: process, and report selected numerical results from an implementation.

18: 

19: \end{abstract}

20: