Fast two- and three-point correlation analysis for time series using spectral methods.
The calculations are FFT-based for optimal performance and offer many options for normalisation, mean removal, averaging, and zero-padding. In particular, averaging over pandas groups of different sizes (e.g. different days) is supported.
| Function | Synopsis |
|---|---|
| acorr | Calculate autocorrelation or autocovariance |
| acorr_grouped_df | Calculate acorr for pandas groups and average |
| corr_mat | Convert correlation vector to matrix |
| fft2x | Calculate cross-bispectrum |
| fftcrop | Return cropped fft or correlation |
| get_nfft | Find a good FFT segment size for pandas groups of different sizes |
| padded_x3corr_norm | Normalise and debias three-point cross-correlations |
| padded_xcorr_norm | Normalise and debias two-point cross-correlations |
| x3corr | Calculate three-point cross-correlation matrix |
| x3corr_grouped_df | Calculate x3corr for pandas groups and average |
| xcorr | Calculate two-point cross-correlation or covariance |
| xcorr_grouped_df | Calculate xcorr for pandas groups and average |
| xcorrshift | Convert xcorr output so lag zero is centered |
The algorithms to calculate three-point correlations and details of daily averaging over high-frequency trading data are described in:
Patzelt, F. and Bouchaud, J-P. (2017): Nonlinear price impact from linear models. Journal of Statistical Mechanics: Theory and Experiment, 12, 123404. Preprint at arXiv:1708.02411.
More code from the same publication is released in the priceprop package.
Please find further explanations in the docstrings and in the examples directory.
pip install scorr
- Python 2.7 or 3.6
- NumPy
- SciPy
- Pandas
- Jupyter
- Matplotlib
- colorednoise