ks_test.py

import scipy.stats
import numpy as np


def ks_test(real_series, simulated_series):
    # `real_series` and `simulated_series` are lists of series, i.e. lists of lists.
    # All series have the same size but we don't need to have as many real series as
    # simulated series.
    real_series, simulated_series = map(np.asarray, (real_series, simulated_series))
    n_steps = len(real_series[0])
    p_values = [0] * n_steps

    for t in range(n_steps):
        real_samples = real_series[:,t]
        simulated_samples = simulated_series[:,t]
        _, p_values[t] = scipy.stats.ks_2samp(real_samples, simulated_samples)

    return p_values


def sliding_ks_test(real_data, simulated_data, sliding_step=1):
    # `real_series` and `simulated_series` are lists of series, i.e. lists of lists.
    # All simulated series have the same length T_sim and all real series have the
    # same length T_real, but T_sim can be different from T_real. The shorter series
    # will be the ones we will make slide over time.
    real_data, simulated_data = map(np.asarray, (real_data, simulated_data))
    real_steps, simulated_steps = real_data.shape[1], simulated_data.shape[1]
    delta_steps = abs(real_steps - simulated_steps)

    if real_steps < simulated_steps:
        data = simulated_data
        pattern = real_data
    else:
        data = real_data
        pattern = simulated_data
    
    p_values = []
    first_step_compared = 0
    while first_step_compared <= delta_steps:
        window = slide(first_step_compared, first_step_compared+pattern.shape[1])
        p_values.append(ks_test(data[:,window], pattern))
        first_step_compared += sliding_step

    return p_values