Nonparametric time series modeling.
Fit multivariate data to a diffusion process.
Alpha stage.
Basically you throw in time series data and out comes the drift vector and noise matrix of a multivariate diffusion process of type
dX_t = f(X_t) dt + g(X_t) dW_t.
Includes a few useful methods, such as a Markov Property check and higher Kramers-Moyal coefficients.
A couple of useful functions are the (Cython optimized) 'crosscorrelate' function, which does cross- correlation (with lag) between two Pandas time series, and 'binner' which estimates the drift and diffusion terms of a (multivariate) diffusion process (also Cythonized... BLAZING fast ;).
The reason I wrote the 'crosscorrelate' function is that scipy.crosscorrelate does a full convolution over the entire dataset (which could be a time series of e.g. 1bn samples), when in practical situations one needs only maybe 100 lag correlation. So this function is therefore a LOT faster!
More stuff coming up soon!