Feature Reference¶
33 features in frozen column order. n = series length, m = mean,
s = population std (ddof=0), d = diff(x). NaN semantics follow
Numerics and NaN policy: min_length is the shortest
defined length, nan_when lists the other undefined conditions.
| # | Feature | Definition | Min length | NaN when | numpy / tsfresh / TSFEL / catch22 |
|---|---|---|---|---|---|
| 0 | mean |
sum(x)/n | 1 | never | np.mean(x) / mean / Mean / — |
| 1 | std |
sqrt(var), population (ddof=0) | 1 | never | np.std(x, ddof=0) / standard_deviation / Standard deviation / — |
| 2 | var |
sum((x-m)^2)/n, forced to exactly 0.0 on constant series | 1 | never | np.var(x, ddof=0) / variance / Variance / — |
| 3 | min |
smallest value | 1 | never | np.min(x) / minimum / Min / — |
| 4 | max |
largest value | 1 | never | np.max(x) / maximum / Max / — |
| 5 | median |
50th percentile, linear interpolation (numpy default) | 1 | never | np.median(x) / median / Median / — |
| 6 | quantile_10 |
10th percentile, linear interpolation | 1 | never | np.percentile(x, 10) / quantile__q_0.1 / — / — |
| 7 | quantile_25 |
25th percentile, linear interpolation | 1 | never | np.percentile(x, 25) / quantile__q_0.25 / — / — |
| 8 | quantile_75 |
75th percentile, linear interpolation | 1 | never | np.percentile(x, 75) / quantile__q_0.75 / — / — |
| 9 | quantile_90 |
90th percentile, linear interpolation | 1 | never | np.percentile(x, 90) / quantile__q_0.9 / — / — |
| 10 | skewness |
(m3/n)/(var*std), m3 = sum((x-m)^3); scipy bias=True | 2 | std == 0 | scipy.stats.skew(x, bias=True) / skewness / Skewness / — |
| 11 | kurtosis |
(m4/n)/var^2 - 3 (Fisher excess); scipy bias=True | 2 | std == 0 | scipy.stats.kurtosis(x, fisher=True, bias=True) / kurtosis / Kurtosis / — |
| 12 | abs_energy |
sum(x^2) | 1 | never | np.sum(x**2) / abs_energy / Absolute energy / — |
| 13 | root_mean_square |
sqrt(sum(x^2)/n) | 1 | never | np.sqrt(np.mean(x**2)) / root_mean_square / Root mean square / — |
| 14 | mean_abs_change |
sum | x[i+1]-x[i] | /(n-1) | 2 |
| 15 | mean_change |
(x[n-1]-x[0])/(n-1) | 2 | n < 2 | np.mean(np.diff(x)) / mean_change / Mean diff / — |
| 16 | cid_ce |
sqrt(sum(((x[i+1]-x[i])/std)^2)); exactly 0.0 when std == 0 | 2 | n < 2 | np.sqrt(np.sum(np.diff(x)**2)) / cid_ce__normalize_False / — / — |
| 17 | mean_second_derivative_central |
sum(x[i+1]-2x[i]+x[i-1])/(2(n-2)) | 3 | n < 3 | np.mean((x[2:] - 2*x[1:-1] + x[:-2]) / 2.0) / mean_second_derivative_central / — / — |
| 18 | zero_crossings |
count of i with (x[i] > 0) != (x[i+1] > 0) | 1 | never | np.sum(np.diff(x > 0).astype(bool)) / — / Zero crossing rate / — |
| 19 | mean_crossings |
count of i with (x[i] > m) != (x[i+1] > m) | 1 | never | np.sum(np.diff(x > np.mean(x)).astype(bool)) / — / — / — |
| 20 | number_of_peaks |
count of i strictly greater than all 3 neighbors on each side | 1 | never (0 when n <= 6) | scipy.signal.find_peaks(x)[0].size / number_peaks__n_1 / Peak to peak distance / — |
| 21 | longest_strike_above_mean |
longest run with x[i] > m | 1 | never | max_streak(x > np.mean(x)) / longest_strike_above_mean / — / — |
| 22 | longest_strike_below_mean |
longest run with x[i] < m | 1 | never | max_streak(x < np.mean(x)) / longest_strike_below_mean / — / — |
| 23 | autocorr_lag_1 |
sum((x[i]-m)(x[i+1]-m))/((n-1)*var) | 2 | n <= 1 or var == 0 | np.corrcoef(x[:-1], x[1:])[0,1] / autocorrelation__lag_1 / Autocorrelation / CO_Embed2_Dist_tau_d_expfit_meandiff |
| 24 | autocorr_lag_2 |
sum((x[i]-m)(x[i+2]-m))/((n-2)*var) | 3 | n <= 2 or var == 0 | np.corrcoef(x[:-2], x[2:])[0,1] / autocorrelation__lag_2 / — / — |
| 25 | autocorr_lag_5 |
sum((x[i]-m)(x[i+5]-m))/((n-5)*var) | 6 | n <= 5 or var == 0 | np.corrcoef(x[:-5], x[5:])[0,1] / autocorrelation__lag_5 / — / — |
| 26 | autocorr_lag_10 |
sum((x[i]-m)(x[i+10]-m))/((n-10)*var) | 11 | n <= 10 or var == 0 | np.corrcoef(x[:-10], x[10:])[0,1] / autocorrelation__lag_10 / — / — |
| 27 | trend_slope |
least-squares slope vs t = 0..n-1; exactly 0.0 on constant series | 2 | n < 2 | scipy.stats.linregress(range(N), x).slope / linear_trend__attr_slope / Linear trend / — |
| 28 | trend_r2 |
squared Pearson r of (t, x) | 2 | n < 2 or var == 0 | scipy.stats.linregress(range(N), x).rvalue**2 / linear_trend__attr_rvalue / — / — |
| 29 | permutation_entropy |
normalized order-3 delay-1 ordinal-pattern entropy (/ln 6) | 3 | n < 3 | antropy.perm_entropy(x, order=3, delay=1, normalize=True) / permutation_entropy__dimension_3__tau_1 / — / — |
| 30 | dominant_frequency |
argmax power-bin frequency (k+1)/n over FFT bins 1..=n/2, DC excluded | 2 | n < 2, constant series, or non-finite/zero total power | np.argmax(np.abs(np.fft.rfft(x))[1:]) + 1 / — / Max power spectrum / SP_Summaries_welch_rect_centroid |
| 31 | spectral_centroid |
sum(f*p)/sum(p) over FFT bins 1..=n/2 | 2 | n < 2, constant series, or non-finite/zero total power | np.sum(freqs * mag) / np.sum(mag) / — / Spectral centroid / SP_Summaries_welch_rect_centroid |
| 32 | spectral_entropy |
shannon(p/sum(p))/ln(nbins); 0.0 when nbins == 1 | 2 | n < 2, constant series, or non-finite/zero total power | antropy.spectral_entropy(x, sf=1.0, method='fft', normalize=True) / — / Spectral entropy / — |
Cost classes: A fused reductions, B order statistics, C spectral/FFT,
D autocorrelations. Intermediates (needs) show shared computation.
| # | Feature | Cost | Needs | Aliases |
|---|---|---|---|---|
| 0 | mean |
A | Needs(PASS1) | tsfresh__mean |
| 1 | std |
A | Needs(PASS1 | PASS2) |
| 2 | var |
A | Needs(PASS1 | PASS2) |
| 3 | min |
A | Needs(PASS1) | minimum, tsfresh__minimum |
| 4 | max |
A | Needs(PASS1) | maximum, tsfresh__maximum |
| 5 | median |
B | Needs(SELECT) | tsfresh__median |
| 6 | quantile_10 |
B | Needs(SELECT) | tsfresh__quantile__q_0.1 |
| 7 | quantile_25 |
B | Needs(SELECT) | tsfresh__quantile__q_0.25 |
| 8 | quantile_75 |
B | Needs(SELECT) | tsfresh__quantile__q_0.75 |
| 9 | quantile_90 |
B | Needs(SELECT) | tsfresh__quantile__q_0.9 |
| 10 | skewness |
A | Needs(PASS1 | PASS2) |
| 11 | kurtosis |
A | Needs(PASS1 | PASS2) |
| 12 | abs_energy |
A | Needs(PASS1) | energy, tsfresh__abs_energy |
| 13 | root_mean_square |
A | Needs(PASS1) | tsfresh__root_mean_square |
| 14 | mean_abs_change |
A | Needs(DIFFS) | tsfresh__mean_abs_change |
| 15 | mean_change |
A | Needs(DIFFS) | tsfresh__mean_change |
| 16 | cid_ce |
A | Needs(PASS1 | PASS2 |
| 17 | mean_second_derivative_central |
A | Needs(DIFFS) | tsfresh__mean_second_derivative_central |
| 18 | zero_crossings |
A | Needs(PASS1 | PASS2) |
| 19 | mean_crossings |
A | Needs(PASS1 | PASS2) |
| 20 | number_of_peaks |
A | Needs(PEAKS) | tsfresh__number_peaks__n_3 |
| 21 | longest_strike_above_mean |
A | Needs(PASS1 | PASS2) |
| 22 | longest_strike_below_mean |
A | Needs(PASS1 | PASS2) |
| 23 | autocorr_lag_1 |
D | Needs(PASS1 | PASS2 |
| 24 | autocorr_lag_2 |
D | Needs(PASS1 | PASS2 |
| 25 | autocorr_lag_5 |
D | Needs(PASS1 | PASS2 |
| 26 | autocorr_lag_10 |
D | Needs(PASS1 | PASS2 |
| 27 | trend_slope |
A | Needs(PASS1 | PASS2 |
| 28 | trend_r2 |
A | Needs(PASS1 | PASS2 |
| 29 | permutation_entropy |
A | Needs(PERM) | tsfresh__permutation_entropy__dimension_3__tau_1 |
| 30 | dominant_frequency |
C | Needs(PASS1 | SPECTRUM) |
| 31 | spectral_centroid |
C | Needs(PASS1 | SPECTRUM) |
| 32 | spectral_entropy |
C | Needs(PASS1 | SPECTRUM) |