tsfresh Cheat Sheet

1Install & the idea0.21

pip install tsfresh★
Pulls in pandas, numpy, scipy, statsmodels. Optional Dask/distributed for scale.
# N time series (by id) -> 1 row of ~1200 features each★
The output is a classic tabular matrix: rows = series, columns = features. Then it's an ordinary ML problem.
from tsfresh import extract_features, select_features, extract_relevant_features★
The three functions you'll use 90% of the time.

2extract_featuresthe core call

X = extract_features(ts, column_id="id", column_sort="time")★
The main call. column_id groups rows into series; column_sort orders each series in time. Returns a wide feature matrix indexed by id.
extract_features(ts, column_id="id", column_sort="time", column_value="value", column_kind="sensor")★
column_kind lets one long table hold multiple signals per id (e.g. several sensors) — features are computed per kind.
extract_features(ts, ..., n_jobs=4, disable_progressbar=True)
Parallelize across cores; extraction is the expensive step (1200 features × N series).

3Data formatget it right first

# LONG (stacked) format — the default input: # columns: id | time | value (one row per observation)★
tsfresh expects long data: each row is one timestamp of one series. Getting this shape right is half the battle.
# multiple signals -> add a "kind" column: # id | time | kind | value
Stack different measurements under a kind column and pass column_kind.
df.melt(id_vars=["id","time"]) # wide -> long, if needed
If your data is wide (one column per series/signal), reshape with pandas melt first.

4Feature settingshow many features

from tsfresh.feature_extraction import EfficientFCParameters extract_features(ts, ..., default_fc_parameters=EfficientFCParameters())★
Efficient skips the most expensive calculators — a good default. ComprehensiveFCParameters() = everything, MinimalFCParameters() = a fast handful.
settings = {"mean": None, "length": None, "fft_coefficient": [{"coeff": 1, "attr": "real"}]}
A custom dict: feature name → parameter list (None = no params). Pass as default_fc_parameters to compute exactly what you want.
# features include: autocorrelation, number_peaks, sample_entropy, # fft_coefficient, cwt_coefficients, ar_coefficient, mean_abs_change...
The catalog spans statistics, spectral, complexity & nonlinearity measures — see the docs' feature list.

5Impute NaNsbefore ML

from tsfresh.utilities.dataframe_functions import impute impute(X) # in place: NaN/inf -> finite values★
Some feature calculators produce NaN/±inf (e.g. on short/constant series). impute replaces them (NaN→median, inf→min/max) so sklearn won't choke.
extract_features(ts, ..., impute_function=impute)★
Impute during extraction in one step.
select_features requires no NaNsorder
Always impute before select_features — the selection test can't run on missing values.

8Reuse settingstrain/serve consistency

from tsfresh.feature_extraction import settings kind_to_fc = settings.from_columns(X_filtered)★
Reconstruct the exact recipe for only the selected features from the filtered matrix's column names.
extract_features(new_ts, ..., kind_to_fc_parameters=kind_to_fc)★
On new/test data, compute only the relevant features — fast, and identical to training. The key to serving.

9Forecasting & sklearnintegrate

from tsfresh.utilities.dataframe_functions import roll_time_series rolled = roll_time_series(ts, column_id="id", column_sort="time", max_timeshift=10, min_timeshift=5)★
Forecasting setup: turn one long series into many rolling windows (each a sub-series id), then extract features per window to predict the next step.
from tsfresh.transformers import RelevantFeatureAugmenter Pipeline([("augmenter", RelevantFeatureAugmenter(...)), ("clf", RandomForestClassifier())])★
A scikit-learn transformer that extracts+selects inside a Pipeline — keeps feature engineering in cross-validation, avoiding leakage.
make_forecasting_frame(series, kind, max_timeshift, rolling_direction)
Helper to build the rolled frame + target for single-series forecasting.

!Common gotchasread before shipping

long format, correct column_id/sort
Wrong id/sort columns silently produce garbage features. Verify the shape (id, time, value) before extracting.
impute before select_features
The selection test errors on NaN/inf. Always impute (or use impute_function=) first.
extraction is expensive
1200 features × many series is slow. Use EfficientFCParameters, n_jobs, or a Dask distributor; drop to a custom settings dict for production.
select needs y; reuse settings for serving
Selection requires labels (training only). At inference use kind_to_fc_parameters from settings.from_columns — don't re-select.
column_kind for multivariate
Multiple signals per id need the kind column, or they'll be mashed into one series.

tsfresh cheat sheet