Dirichlet Process Mixtures

A standard mixture model, like a Gaussian mixture, asks you to decide the number of clusters up front — three regimes, five customer segments — and then fits the data to that fixed count. A Dirichlet process mixture removes that requirement: it treats the number of clusters itself as something to be learned, allowing new clusters to appear as more data arrives if the existing ones no longer fit well. A useful way to picture it is the "Chinese restaurant process": each new observation either joins an existing table (cluster) with probability proportional to how many points are already there, or starts a brand-new table with a probability controlled by a single concentration parameter — a small value favors a few large clusters, a large value favors many small ones. As data accumulates, popular clusters attract more new points while rarely-used ones stay small or disappear from active consideration, and the model never has to commit to a cluster count that later turns out to be wrong.

This is useful whenever the natural number of regimes, strategies, or subpopulations in a dataset is genuinely unknown and might grow — for example, discovering an unanticipated new market regime without having pre-specified how many regimes exist.