A library for SciKit-Inspired Time Series models.
The primary goal of this library is to allow one to train time series prediction models using a similar API to
scikit-learn. Consequently, similar to
scikit-learn, this library consists of
Clone the library, create a virtual environment, and install the dependencies.
Do this in one fell swoop with conda
conda env create -f environment.yml -n skits
Or install with pip after creating a virtual environment
pip install -r requirements.txt
You can install the library locally with
pip install -e .
The preprocessors expect to receive time series data, and then end up storing some data about the time series such that they can fully invert a transform. The following example shows how to create a
DifferenceTransformer transform data, and then invert it back to its original form. The
DifferenceTransformer subtracts the point shifted by
period away from each point.
import numpy as np from skits.preprocessing import DifferenceTransformer y = np.random.random(10) # scikit-learn expects 2D design matrices, # so we duplicate the time series. X = y[:, np.newaxis] dt = DifferenceTransformer(period=2) Xt = dt.fit_transform(X,y) X_inv = dt.inverse_transform(Xt) assert np.allclose(X, X_inv)
After all preprocessing transformations are completed, multiple features may be built out of the time series. These can be built via feature extractors, which one should combine together into a large
FeatureUnion. Current features include autoregressive, seasonal, and integrated features (covering the AR and I of ARIMA models).
There are two types of pipelines. The
ForecasterPipeline is for forecasting time series (duh). Specifically, one should build this pipeline with a regressor as the final step such that one can make appropriate predictions. The functionality is similar to a regular
scikit-learn pipeline. Differences include the addition of a
forecast() method along with a
to_scale keyword argument to
predict() such that one can make sure that their prediction is on the same scale as the original data.
These classes are likely subject to change as they are fairly hacky right now. For example, one must transform both
y for all transformations before the introduction of a
DifferenceTransformer. While the pipeline handles this, one must prefix all of these transformations with
pre_ in the step names.
Anywho, here's an example:
import numpy as np from sklearn.linear_model import LinearRegression from sklearn.preprocessing import StandardScaler from sklearn.pipeline import FeatureUnion from skits.pipeline import ForecasterPipeline from skits.preprocessing import ReversibleImputer from skits.feature_extraction import (AutoregressiveTransformer, SeasonalTransformer) steps = [ ('pre_scaling', StandardScaler()), ('features', FeatureUnion([ ('ar_transformer', AutoregressiveTransformer(num_lags=3)), ('seasonal_transformer', SeasonalTransformer(seasonal_period=20) )])), ('post_features_imputer', ReversibleImputer()), ('regressor', LinearRegression(fit_intercept=False)) ] l = np.linspace(0, 1, 101) y = 5*np.sin(2 * np.pi * 5 * l) + np.random.normal(0, 1, size=101) X = y[:, np.newaxis] pipeline = ForecasterPipeline(steps) pipeline.fit(X, y) y_pred = pipeline.predict(X, to_scale=True, refit=True)
And this ends up looking like:
import matplotlib.pyplot as plt plt.plot(y, lw=2) plt.plot(y_pred, lw=2) plt.legend(['y_true', 'y_pred'], bbox_to_anchor=(1, 1));
And forecasting looks like
start_idx = 70 plt.plot(y, lw=2); plt.plot(pipeline.forecast(y[:, np.newaxis], start_idx=start_idx), lw=2); ax = plt.gca(); ylim = ax.get_ylim(); plt.plot((start_idx, start_idx), ylim, lw=4); plt.ylim(ylim); plt.legend(['y_true', 'y_pred', 'forecast start'], bbox_to_anchor=(1, 1));