A useful machine-learning application would be to find an "executive summary" of a time series. By this I mean another time series, which when subtracted from the original series yields a series that can be captured by a simple statistical model (e.g. a normal variate). The utility of such a "tendency" is that it would encapsulate the most critical aspects a dynamicist might need to explain if this signal has a rational basis. The intrinsic time decomposition is shown to be a diffusion filter that can handle multiscale dynamics. A "tendency" can be produced from an intrinsic time decomposition, followed by a selection criterion that favors empirical statistical symmetries and information content. The tendency, when applied to a climate signal, will show the warming trend as well as the preponderance of the ocean in controlling the time scales of fluctuations.