hctsa
hctsa is a comprehensive package of thousands of time-series analysis methods. The software allows the user to convert a time series into a vector of thousands of informative features, corresponding to different outputs of time-series analysis operations. Converting from a time series to a feature vector is useful for many tasks; for example, we have used it to tackle the following problems:
- Classifying seizures from EEG brain recordings
- Diagnosing Parkinson's disease from speech for precision medicine
- Monitoring sleep-stage progression
- Predicting schizophrenia from brain-imaging data
- Forecasting catastrohpes in financial and ecological systems
- General time-series classification problems (see some umap projections here)
Time-series analysis features
The types of features included are diverse, and range from:
- Linear correlation (autocorrelation, Fourier spectral density)
- Information theoretic quantities (Approximate/Sample/Permutation entropy, automutual information)
- Model fits (AR, ARMA, GARCH, Gaussian process regression, exponential smoothing, state space, HMM)
- Nonlinear time-series analysis (fractal dimension estimates)
- Scaling (detrended fluctuation analysis)
- Stationarity (sliding window measures, StatAv)
- Properties of the distribution (skewness, symmetry, distribution type)
- Basis functions (wavelets)
- Other measures (extreme events, visibility graphs, delay-vector variance)