Time Series

Data scientists spend much of their time cleaning data or building features. While the former is unavoidable, the latter can be automated.

tsfresh uses a robust feature selection algorithm to automatically extract relevant time series features, freeing up data scientists’ time.

To demonstrate this, start with loading an example dataset:

from tsfresh.examples.robot_execution_failures import (
download_robot_execution_failures,
load_robot_execution_failures,
)

download_robot_execution_failures()
timeseries, y = load_robot_execution_failures()
timeseries.head()

Output:

id time F_x F_y F_z T_x T_y T_z
0 1 0 -1 -1 63 -3 -1 0
1 1 1 0 0 62 -3 -1 0
2 1 2 -1 -1 61 -3 0 0
3 1 3 -1 -1 63 -2 -1 0
4 1 4 -1 -1 63 -3 -1 0

Extract features and select only relevant features for each time series.

from tsfresh import extract_relevant_features

# extract relevant features
features_filtered = extract_relevant_features(
timeseries, y, column_id="id", column_sort="time"
)

You can now use the features in features_filtered to train your classification model.

Link to tsfresh.

Interact with this code in Google Colab.

Traditional time series analysis separates preprocessing tasks, such as computing lags, applying transformations, and feature engineering, from the actual model fitting process, making the overall workflow more cumbersome.

In contrast, MLForecast integrates preprocessing tasks within a single class. This streamlines the workflow, enabling easy experimentation with different preprocessing and modeling combinations.

Link to MLForecast

Traditional time series libraries face limitations with large datasets as they run in memory on single machines.

StatsForecast overcomes this by providing seamless Spark compatibility, enabling scalable and efficient time-series forecasting on big data directly in Spark.

Link to StatsForecast.

TimeGPT is a powerful generative pre-trained forecasting model that can generate accurate forecasts for new time series without the need for training.

With TimeGPT, you can accurately predict various domains such as retail, electricity, finance, and IoT with just a few lines of code.

Link to TimeGPT.

Neural forecasting methods can enhance forecasting accuracy, but they are often difficult to use and computationally expensive.

NeuralForecast provides a simple way to use efficient models, using familiar scikit-learn syntax. The models available in NeuralForecast range from classic networks like RNN to the latest transformers.

Link to NeuralForecast.

Knee-point detection in time series identifies the point of maximum curvature. The knee point can identify anomalies or outliers in the time series. If a data point is far away from the knee point, it may indicate an anomaly or unexpected behavior.

The kneed library makes it easy to implement knee-point detection in Python.

Link to kneed.

You’re running a marketing campaign and see a user increase. But how do you know if it’s due to the campaign or just a coincidence?

That is when tfcausalimpact comes in handy. It forecasts future data trends using a Bayesian structural model and compares them to actual data to extract statistical conclusions.

Link to tfcausalimpact.

Probabilistic models offer a range of possible future outcomes, rather than a single fixed prediction, allowing for the assessment of risk associated with adverse events. 

GluonTS streamlines the process of using probabilistic models for time series data.

Link to GluonTS.