Feature Extraction

tseda.features

Feature extraction for time series.

Public API

TemporalFeatureExtractor: Calendar and cyclic time-index features → pandas.DataFrame.
StatisticalFeatureExtractor: Distribution, complexity, and linear-structure features → single-row DataFrame.
SpectralFeatureExtractor: Frequency-domain (FFT) features → single-row DataFrame.

class tseda.features.TemporalFeatureExtractor[source]

Bases: object

Extract calendar and cyclic time features from a TimeSeries.

extract(ts, cyclic, time_index)[source]

Return a pandas.DataFrame with one feature column per row aligned to ts.index.

Parameters:

ts (TimeSeries)
cyclic (bool)
time_index (bool)

Return type:

DataFrame

Examples

>>> import pandas as pd, numpy as np
>>> from tseda import TimeSeries
>>> from tseda.features.temporal import TemporalFeatureExtractor

>>> idx = pd.date_range("2020-01-01", periods=5, freq="D")
>>> ts  = TimeSeries([10.0, 11.0, 12.0, 11.5, 10.5], index=idx)
>>> df  = TemporalFeatureExtractor().extract(ts)
>>> int(df["year"].iloc[0])
2020
>>> int(df["month"].iloc[0])
1

extract(ts, *, cyclic=True, time_index=True)[source]

Extract temporal features aligned to ts.index.

Parameters:

ts (TimeSeries) – Input series.
cyclic (bool, optional) – When True (default), add sine/cosine encodings for month, dayofweek, and hour.
time_index (bool, optional) – When True (default), add days_since_start and time_norm (0 → 1 over the series span).

Returns:

Index matches ts.index. Columns:

Always present:: year, month, day, dayofweek, hour, quarter, weekofyear, is_weekend, is_month_start, is_month_end.
When cyclic=True:: month_sin, month_cos, dow_sin, dow_cos, hour_sin, hour_cos.
When time_index=True:: days_since_start, time_norm.

Return type:

pandas.DataFrame

Raises:

TypeError – If ts is not a TimeSeries.

Examples

>>> import pandas as pd, numpy as np
>>> from tseda import TimeSeries
>>> from tseda.features.temporal import TemporalFeatureExtractor
>>> idx = pd.date_range("2020-01-01", periods=7, freq="D")
>>> ts  = TimeSeries(np.ones(7), index=idx)
>>> df  = TemporalFeatureExtractor().extract(ts, cyclic=False, time_index=False)
>>> set(df.columns) >= {"year", "month", "day", "dayofweek", "is_weekend"}
True

class tseda.features.StatisticalFeatureExtractor[source]

Bases: object

Extract statistical features from a TimeSeries.

The extractor is stateless. It operates on the non-NaN values of the series.

extract(ts, entropy)[source]

Return a single-row pandas.DataFrame of features.

Parameters:

ts (TimeSeries)
entropy (bool)

Return type:

DataFrame

Examples

>>> import pandas as pd, numpy as np
>>> from tseda import TimeSeries
>>> from tseda.features.statistical import StatisticalFeatureExtractor

>>> idx = pd.date_range("2020", periods=100, freq="D")
>>> ts  = TimeSeries(np.arange(100.0), index=idx)
>>> df  = StatisticalFeatureExtractor().extract(ts)
>>> round(float(df["linear_slope"].iloc[0]), 1)
1.0
>>> round(float(df["linear_r2"].iloc[0]), 2)
1.0

extract(ts, *, entropy=True)[source]

Compute statistical features for ts.

Parameters:

ts (TimeSeries) – Input series.
entropy (bool, optional) – When True (default), compute approximate entropy and sample entropy. These are O(n²) — set False for large series (n > 2000) to save time.

Returns:

One row, columns:

Distribution:: mean, std, var, skewness, kurtosis, min, max, range, median, iqr, mad, cv, trimmed_mean, q25, q75, q05, q95.
Complexity:: turning_points_ratio, mean_crossing_rate, flatness_ratio. If entropy=True: approx_entropy, sample_entropy.
Linear structure:: lag1_acf, linear_slope, linear_r2.
Nonlinearity:: n_peaks, n_troughs.

Return type:

pandas.DataFrame

Raises:

TypeError – If ts is not a TimeSeries.
ValueError – If ts has fewer than 4 non-NaN observations.

Examples

>>> import pandas as pd, numpy as np
>>> from tseda import TimeSeries
>>> from tseda.features.statistical import StatisticalFeatureExtractor
>>> idx = pd.date_range("2020", periods=5, freq="D")
>>> ts  = TimeSeries([1.0, 2.0, 1.5, 2.5, 2.0], index=idx)
>>> df  = StatisticalFeatureExtractor().extract(ts, entropy=False)
>>> float(df["mean"].iloc[0])
1.8

class tseda.features.SpectralFeatureExtractor[source]

Bases: object

Extract frequency-domain features from a TimeSeries.

The extractor is stateless. NaN values in the series are replaced by linear interpolation before FFT analysis.

extract(ts, n_bands)[source]

Return a single-row pandas.DataFrame of spectral features.

Parameters:

ts (TimeSeries)
n_bands (int)

Return type:

DataFrame

Examples

>>> import pandas as pd, numpy as np
>>> from tseda import TimeSeries
>>> from tseda.features.spectral import SpectralFeatureExtractor

>>> idx = pd.date_range("2020", periods=128, freq="h")
>>> ts  = TimeSeries(np.sin(2 * np.pi * np.arange(128) / 24), index=idx)
>>> df  = SpectralFeatureExtractor().extract(ts)
>>> "spectral_centroid" in df.columns
True

extract(ts, *, n_bands=3)[source]

Compute spectral features for ts.

Parameters:

ts (TimeSeries) – Input series. NaN values are linearly interpolated before the FFT.
n_bands (int, optional) – Number of equal-width frequency bands for band power features. Default 3 (low / mid / high). Must be >= 1.

Returns:

One row, columns:

Energy:: total_power, band_power_0 … band_power_{n_bands-1}.
Shape:: spectral_centroid, spectral_bandwidth, spectral_rolloff_0.5, spectral_rolloff_0.85, spectral_entropy, spectral_flatness.
Peak:: dominant_freq, dominant_period, n_spectral_peaks.

Return type:

pandas.DataFrame

Raises:

TypeError – If ts is not a TimeSeries.
ValueError – If the series has fewer than 8 non-NaN observations or n_bands < 1.

Examples

>>> import pandas as pd, numpy as np
>>> from tseda import TimeSeries
>>> from tseda.features.spectral import SpectralFeatureExtractor
>>> idx = pd.date_range("2020", periods=256, freq="D")
>>> ts  = TimeSeries(np.cos(2*np.pi*np.arange(256)/7), index=idx)
>>> df  = SpectralFeatureExtractor().extract(ts)
>>> int(round(float(df["dominant_period"].iloc[0])))
7

Temporal feature extraction for time series.

Extracts calendar-based and time-index features from a TimeSeries. All features are deterministic functions of the datetime index — no statistical estimation required.

Two categories are produced:

Calendar features — year, month, day, hour, day-of-week, quarter, and boolean flags (is_weekend, is_month_start, is_month_end).
Cyclic encodings — sine/cosine projections of periodic calendar fields (month, day-of-week, hour) so that month 12 and month 1 are close in feature space.

Classes

TemporalFeatureExtractor: Stateless extractor returning a pandas.DataFrame.

Examples

>>> import pandas as pd, numpy as np
>>> from tseda import TimeSeries
>>> from tseda.features.temporal import TemporalFeatureExtractor

>>> idx = pd.date_range("2020-01-01", periods=10, freq="D")
>>> ts  = TimeSeries(np.arange(10.0), index=idx)
>>> df  = TemporalFeatureExtractor().extract(ts)
>>> list(df.columns[:4])
['year', 'month', 'day', 'dayofweek']

class tseda.features.temporal.TemporalFeatureExtractor[source]

Bases: object

Extract calendar and cyclic time features from a TimeSeries.

extract(ts, cyclic, time_index)[source]

Return a pandas.DataFrame with one feature column per row aligned to ts.index.

Parameters:

ts (TimeSeries)
cyclic (bool)
time_index (bool)

Return type:

DataFrame

Examples

>>> import pandas as pd, numpy as np
>>> from tseda import TimeSeries
>>> from tseda.features.temporal import TemporalFeatureExtractor

>>> idx = pd.date_range("2020-01-01", periods=5, freq="D")
>>> ts  = TimeSeries([10.0, 11.0, 12.0, 11.5, 10.5], index=idx)
>>> df  = TemporalFeatureExtractor().extract(ts)
>>> int(df["year"].iloc[0])
2020
>>> int(df["month"].iloc[0])
1

extract(ts, *, cyclic=True, time_index=True)[source]

Extract temporal features aligned to ts.index.

Parameters:

ts (TimeSeries) – Input series.
cyclic (bool, optional) – When True (default), add sine/cosine encodings for month, dayofweek, and hour.
time_index (bool, optional) – When True (default), add days_since_start and time_norm (0 → 1 over the series span).

Returns:

Index matches ts.index. Columns:

Always present:: year, month, day, dayofweek, hour, quarter, weekofyear, is_weekend, is_month_start, is_month_end.
When cyclic=True:: month_sin, month_cos, dow_sin, dow_cos, hour_sin, hour_cos.
When time_index=True:: days_since_start, time_norm.

Return type:

pandas.DataFrame

Raises:

TypeError – If ts is not a TimeSeries.

Examples

>>> import pandas as pd, numpy as np
>>> from tseda import TimeSeries
>>> from tseda.features.temporal import TemporalFeatureExtractor
>>> idx = pd.date_range("2020-01-01", periods=7, freq="D")
>>> ts  = TimeSeries(np.ones(7), index=idx)
>>> df  = TemporalFeatureExtractor().extract(ts, cyclic=False, time_index=False)
>>> set(df.columns) >= {"year", "month", "day", "dayofweek", "is_weekend"}
True

Statistical feature extraction for time series.

Extracts a rich set of statistical descriptors that characterise the distribution, complexity, and structure of a TimeSeries. All features are computed in pure numpy.

Feature groups

Distribution — mean, std, skewness, kurtosis, quantiles, range, CV.
Spread / Robust — MAD, trimmed mean, IQR.
Complexity — approximate entropy, sample entropy, turning points ratio, mean-crossing rate.
Linear structure — lag-1 autocorrelation, linear-trend slope and R².
Nonlinearity — number of peaks and troughs, flatness ratio.

Classes

StatisticalFeatureExtractor: Stateless extractor.

Examples

>>> import pandas as pd, numpy as np
>>> from tseda import TimeSeries
>>> from tseda.features.statistical import StatisticalFeatureExtractor

>>> rng = np.random.default_rng(0)
>>> idx = pd.date_range("2020", periods=200, freq="D")
>>> ts  = TimeSeries(rng.standard_normal(200), index=idx)
>>> df  = StatisticalFeatureExtractor().extract(ts)
>>> "mean" in df.columns and "std" in df.columns
True
>>> df.shape[0]
1

class tseda.features.statistical.StatisticalFeatureExtractor[source]

Bases: object

Extract statistical features from a TimeSeries.

The extractor is stateless. It operates on the non-NaN values of the series.

extract(ts, entropy)[source]

Return a single-row pandas.DataFrame of features.

Parameters:

ts (TimeSeries)
entropy (bool)

Return type:

DataFrame

Examples

>>> import pandas as pd, numpy as np
>>> from tseda import TimeSeries
>>> from tseda.features.statistical import StatisticalFeatureExtractor

>>> idx = pd.date_range("2020", periods=100, freq="D")
>>> ts  = TimeSeries(np.arange(100.0), index=idx)
>>> df  = StatisticalFeatureExtractor().extract(ts)
>>> round(float(df["linear_slope"].iloc[0]), 1)
1.0
>>> round(float(df["linear_r2"].iloc[0]), 2)
1.0

extract(ts, *, entropy=True)[source]

Compute statistical features for ts.

Parameters:

ts (TimeSeries) – Input series.
entropy (bool, optional) – When True (default), compute approximate entropy and sample entropy. These are O(n²) — set False for large series (n > 2000) to save time.

Returns:

One row, columns:

Distribution:: mean, std, var, skewness, kurtosis, min, max, range, median, iqr, mad, cv, trimmed_mean, q25, q75, q05, q95.
Complexity:: turning_points_ratio, mean_crossing_rate, flatness_ratio. If entropy=True: approx_entropy, sample_entropy.
Linear structure:: lag1_acf, linear_slope, linear_r2.
Nonlinearity:: n_peaks, n_troughs.

Return type:

pandas.DataFrame

Raises:

TypeError – If ts is not a TimeSeries.
ValueError – If ts has fewer than 4 non-NaN observations.

Examples

>>> import pandas as pd, numpy as np
>>> from tseda import TimeSeries
>>> from tseda.features.statistical import StatisticalFeatureExtractor
>>> idx = pd.date_range("2020", periods=5, freq="D")
>>> ts  = TimeSeries([1.0, 2.0, 1.5, 2.5, 2.0], index=idx)
>>> df  = StatisticalFeatureExtractor().extract(ts, entropy=False)
>>> float(df["mean"].iloc[0])
1.8

Spectral feature extraction for time series.

Computes frequency-domain descriptors of a TimeSeries using the FFT power spectrum. All computations use pure numpy / scipy.

Feature groups

Energy / Power — total spectral power, power in low / mid / high bands.
Shape — spectral centroid, bandwidth, rolloff frequency, spectral entropy.
Peak — dominant frequency, dominant period, number of spectral peaks.
Temporal — spectral flatness (ratio of geometric to arithmetic mean power).

Classes

SpectralFeatureExtractor: Stateless extractor.

Examples

>>> import pandas as pd, numpy as np
>>> from tseda import TimeSeries
>>> from tseda.features.spectral import SpectralFeatureExtractor

>>> rng = np.random.default_rng(0)
>>> idx = pd.date_range("2020", periods=256, freq="D")
>>> ts  = TimeSeries(np.sin(2 * np.pi * np.arange(256) / 7), index=idx)
>>> df  = SpectralFeatureExtractor().extract(ts)
>>> int(round(float(df["dominant_period"].iloc[0])))
7

class tseda.features.spectral.SpectralFeatureExtractor[source]

Bases: object

Extract frequency-domain features from a TimeSeries.

The extractor is stateless. NaN values in the series are replaced by linear interpolation before FFT analysis.

extract(ts, n_bands)[source]

Return a single-row pandas.DataFrame of spectral features.

Parameters:

ts (TimeSeries)
n_bands (int)

Return type:

DataFrame

Examples

>>> import pandas as pd, numpy as np
>>> from tseda import TimeSeries
>>> from tseda.features.spectral import SpectralFeatureExtractor

>>> idx = pd.date_range("2020", periods=128, freq="h")
>>> ts  = TimeSeries(np.sin(2 * np.pi * np.arange(128) / 24), index=idx)
>>> df  = SpectralFeatureExtractor().extract(ts)
>>> "spectral_centroid" in df.columns
True

extract(ts, *, n_bands=3)[source]

Compute spectral features for ts.

Parameters:

ts (TimeSeries) – Input series. NaN values are linearly interpolated before the FFT.
n_bands (int, optional) – Number of equal-width frequency bands for band power features. Default 3 (low / mid / high). Must be >= 1.

Returns:

One row, columns:

Energy:: total_power, band_power_0 … band_power_{n_bands-1}.
Shape:: spectral_centroid, spectral_bandwidth, spectral_rolloff_0.5, spectral_rolloff_0.85, spectral_entropy, spectral_flatness.
Peak:: dominant_freq, dominant_period, n_spectral_peaks.

Return type:

pandas.DataFrame

Raises:

TypeError – If ts is not a TimeSeries.
ValueError – If the series has fewer than 8 non-NaN observations or n_bands < 1.

Examples

>>> import pandas as pd, numpy as np
>>> from tseda import TimeSeries
>>> from tseda.features.spectral import SpectralFeatureExtractor
>>> idx = pd.date_range("2020", periods=256, freq="D")
>>> ts  = TimeSeries(np.cos(2*np.pi*np.arange(256)/7), index=idx)
>>> df  = SpectralFeatureExtractor().extract(ts)
>>> int(round(float(df["dominant_period"].iloc[0])))
7