Source code for statsmodels.tsa._bds

"""
BDS test for IID time series

References
----------

Broock, W. A., J. A. Scheinkman, W. D. Dechert, and B. LeBaron. 1996.
"A Test for Independence Based on the Correlation Dimension."
Econometric Reviews 15 (3): 197-235.

Kanzler, Ludwig. 1999.
"Very Fast and Correctly Sized Estimation of the BDS Statistic".
SSRN Scholarly Paper ID 151669. Rochester, NY: Social Science Research Network.

LeBaron, Blake. 1997.
"A Fast Algorithm for the BDS Statistic."
Studies in Nonlinear Dynamics & Econometrics 2 (2) (January 1).
"""

from __future__ import division
import numpy as np
from scipy import stats


def distance_indicators(x, epsilon=None, distance=1.5):
    """
    Calculate all pairwise threshold distance indicators for a time series

    Parameters
    ----------
    x : 1d array
        observations of time series for which heaviside distance indicators
        are calculated
    epsilon : scalar, optional
        the threshold distance to use in calculating the heaviside indicators
    distance : scalar, optional
        if epsilon is omitted, specifies the distance multiplier to use when
        computing it

    Returns
    -------
    indicators : 2d array
        matrix of distance threshold indicators

    Notes
    -----
    Since this can be a very large matrix, use np.int8 to save some space.

    """
    x = np.asarray(x)
    nobs = len(x)

    if epsilon is not None and epsilon <= 0:
        raise ValueError("Threshold distance must be positive if specified."
                         " Got epsilon of %f" % epsilon)
    if distance <= 0:
        raise ValueError("Threshold distance must be positive."
                         " Got distance multiplier %f" % distance)

    # TODO: add functionality to select epsilon optimally
    # TODO: and/or compute for a range of epsilons in [0.5*s, 2.0*s]?
    #      or [1.5*s, 2.0*s]?
    if epsilon is None:
        epsilon = distance * x.std(ddof=1)

    return np.abs(x[:, None] - x) < epsilon


def correlation_sum(indicators, embedding_dim):
    """
    Calculate a correlation sum

    Useful as an estimator of a correlation integral

    Parameters
    ----------
    indicators : 2d array
        matrix of distance threshold indicators
    embedding_dim : integer
        embedding dimension

    Returns
    -------
    corrsum : float
        Correlation sum
    indicators_joint
        matrix of joint-distance-threshold indicators

    """
    if not indicators.ndim == 2:
        raise ValueError('Indicators must be a matrix')
    if not indicators.shape[0] == indicators.shape[1]:
        raise ValueError('Indicator matrix must be symmetric (square)')

    if embedding_dim == 1:
        indicators_joint = indicators
    else:
        corrsum, indicators = correlation_sum(indicators, embedding_dim - 1)
        indicators_joint = indicators[1:, 1:]*indicators[:-1, :-1]

    nobs = len(indicators_joint)
    corrsum = np.mean(indicators_joint[np.triu_indices(nobs, 1)])
    return corrsum, indicators_joint


def correlation_sums(indicators, max_dim):
    """
    Calculate all correlation sums for embedding dimensions 1:max_dim

    Parameters
    ----------
    indicators : 2d array
        matrix of distance threshold indicators
    max_dim : integer
        maximum embedding dimension

    Returns
    -------
    corrsums : 1d array
        Correlation sums

    """

    corrsums = np.zeros((1, max_dim))

    corrsums[0, 0], indicators = correlation_sum(indicators, 1)
    for i in range(1, max_dim):
        corrsums[0, i], indicators = correlation_sum(indicators, 2)

    return corrsums


def _var(indicators, max_dim):
    """
    Calculate the variance of a BDS effect

    Parameters
    ----------
    indicators : 2d array
        matrix of distance threshold indicators
    max_dim : integer
        maximum embedding dimension

    Returns
    -------
    variances : float
        Variance of BDS effect

    """
    nobs = len(indicators)
    corrsum_1dim, _ = correlation_sum(indicators, 1)
    k = ((indicators.sum(1)**2).sum() - 3*indicators.sum() +
         2*nobs) / (nobs * (nobs - 1) * (nobs - 2))

    variances = np.zeros((1, max_dim - 1))

    for embedding_dim in range(2, max_dim + 1):
        tmp = 0
        for j in range(1, embedding_dim):
            tmp += (k**(embedding_dim - j))*(corrsum_1dim**(2 * j))
        variances[0, embedding_dim-2] = 4 * (
            k**embedding_dim +
            2 * tmp +
            ((embedding_dim - 1)**2) * (corrsum_1dim**(2 * embedding_dim)) -
            (embedding_dim**2) * k * (corrsum_1dim**(2 * embedding_dim - 2)))

    return variances, k


[docs]def bds(x, max_dim=2, epsilon=None, distance=1.5): """ Calculate the BDS test statistic for independence of a time series Parameters ---------- x : 1d array observations of time series for which bds statistics is calculated max_dim : integer maximum embedding dimension epsilon : scalar, optional the threshold distance to use in calculating the correlation sum distance : scalar, optional if epsilon is omitted, specifies the distance multiplier to use when computing it Returns ------- bds_stat : float The BDS statistic pvalue : float The p-values associated with the BDS statistic Notes ----- The null hypothesis of the test statistic is for an independent and identically distributed (i.i.d.) time series, and an unspecified alternative hypothesis. This test is often used as a residual diagnostic. The calculation involves matrices of size (nobs, nobs), so this test will not work with very long datasets. Implementation conditions on the first m-1 initial values, which are required to calculate the m-histories: x_t^m = (x_t, x_{t-1}, ... x_{t-(m-1)}) """ x = np.asarray(x) nobs_full = len(x) if max_dim < 2 or max_dim >= nobs_full: raise ValueError("Maximum embedding dimension must be in the range" " [2,len(x)-1]. Got %d." % max_dim) # Cache the indicators indicators = distance_indicators(x, epsilon, distance) # Get estimates of m-dimensional correlation integrals corrsum_mdims = correlation_sums(indicators, max_dim) # Get variance of effect variances, k = _var(indicators, max_dim) stddevs = np.sqrt(variances) bds_stats = np.zeros((1, max_dim - 1)) pvalues = np.zeros((1, max_dim - 1)) for embedding_dim in range(2, max_dim+1): ninitial = (embedding_dim - 1) nobs = nobs_full - ninitial # Get estimates of 1-dimensional correlation integrals # (see Kanzler footnote 10 for why indicators are truncated) corrsum_1dim, _ = correlation_sum(indicators[ninitial:, ninitial:], 1) corrsum_mdim = corrsum_mdims[0, embedding_dim - 1] # Get the intermediate values for the statistic effect = corrsum_mdim - (corrsum_1dim**embedding_dim) sd = stddevs[0, embedding_dim - 2] # Calculate the statistic: bds_stat ~ N(0,1) bds_stats[0, embedding_dim - 2] = np.sqrt(nobs) * effect / sd # Calculate the p-value (two-tailed test) pvalue = 2*stats.norm.sf(np.abs(bds_stats[0, embedding_dim - 2])) pvalues[0, embedding_dim - 2] = pvalue return np.squeeze(bds_stats), np.squeeze(pvalues)