Python AnalySis for Time Series

This package aims to structure the way time series analysis and forecasting is done.

Purpose of the Package

• The purpose of the package is to provide a collection of forecasting models and analysis methods for time series in one unified library.

Features

• Collection of analysis methods:
  • Scipy and statsmodel for testing
  • Time series processing
  • Statistical testing (stationarity, check seasonality, ... )
  • Visualization
• Collection of forecasting models using Darts, which is itself an aggregator of
  • scikit-learn
  • tenserflow
  • prophet
  • Auto regression models (ARIMA, SARIMA, ....)
  • etc.

Installation

The package can be installed by :

python3 -m pip install git+https://github.com/eurobios-mews-labs/pasts

Building the documentation

First, make sure you have sphinx and the Readthedocs theme installed.

If you use pip, open a terminal and enter the following commands:

pip install sphinx
pip install sphinx_rtd_theme

If you use conda, open an Anaconda Powershell Prompt and enter the following commands:

conda install sphinx
conda install sphinx_rtd_theme

Then, in the same terminal or anaconda prompt, build the doc with:

cd doc
make html

The documentation can then be accessed from doc/_build/html/index.html.

Usage and example

You can find examples for the Signal class for univariate and multivariate series here: examples/ex_model.py

The Operation class can be used on its own. Find an example here: examples/ex_operations.py

Start project

To start using the package, import you data as a pandas dataframe with a temporal index and use the Signal class.

import pandas as pd

from darts.datasets import AirPassengersDataset
from darts.models import AutoARIMA, Prophet, ExponentialSmoothing, XGBModel, VARIMA
from darts.utils.utils import ModelMode, SeasonalityMode

from pasts.signal import Signal
from pasts.visualization import Visualization

series = AirPassengersDataset().load()
dt = pd.DataFrame(series.values())
dt.rename(columns={0: 'passengers'}, inplace=True)
dt.index = series.time_index
signal = Signal(dt)

Visualize and analyze data

The properties attribute contains some information about the data. Use the Visualization class to generate various types of plots.

print(signal.properties)

Output:

>>> {'shape': (144, 1), 'types': passengers    float64
dtype: object, 
'is_univariate': True, 
'nanSum': passengers   0
dtype: int64, 
'quantiles':   0.00   0.05   0.50    0.95    0.99   1.00
passengers     104.0  121.6  265.5  488.15  585.79  622.0}

Visualization(signal).plot_signal()
Visualization(signal).acf_plot()

Yield:

You can also perform some statistical tests specific to time series.

signal.apply_stat_test('stationary')
signal.apply_stat_test('stationary', 'kpss')
signal.apply_stat_test('seasonality')
print(signal.tests_stat)

Output: Whether the null hypothesis is rejected and p-value for stationarity, seasonality period for seasonality.

>>> {'stationary: adfuller': (False, 0.9918802434376409),
'stationary: kpss': (False, 0.01),
'seasonality: check_seasonality': (<function check_seasonality at 0x000001D1D62EE310>, (True, 12))}

Machine Learning

Choose a date to split the series between train and test.

timestamp = '1958-12-01'
signal.validation_split(timestamp=timestamp)

The library provides some operations to apply before using forecasting models. In this example, both linear trend and seasonality are removed. Machine Learning models will be trained on the remainig series, then inverse functions will be applied to the predicted series.

signal.apply_operations(['trend', 'seasonality'])
Visualization(signal).plot_signal()

Use the method apply_model to apply models of your choice. If the parameters gridsearch and parameters are passed, a gridsearch will be performed.

signal.apply_model(ExponentialSmoothing())

signal.apply_model(AutoARIMA())
signal.apply_model(Prophet())

# Be careful : if trend and seasonality were removed, this specific gridsearch cannot be performed.
param_grid = {'trend': [ModelMode.ADDITIVE, ModelMode.MULTIPLICATIVE, ModelMode.NONE],
              'seasonal': [SeasonalityMode.ADDITIVE, SeasonalityMode.MULTIPLICATIVE, SeasonalityMode.NONE],
              }
signal.apply_model(ExponentialSmoothing(), gridsearch=True, parameters=param_grid)

You can pass a list of metrics to the method compute_scores. By default, it tries to compute R², MSE, RMSE, MAPE, SMAPE and MAE. Warnings are raised if some metrics are impossible to compute with this type of data. You can choose to compute scores time-wise or unit-wise with parameter axis. However, with univariate data it is preferable to keep the default value (axis=1, unit-wise).

signal.compute_scores(list_metrics=['rmse', 'r2'])
signal.compute_scores()
print(signal.models['Prophet']['scores'])

Output:

>>> {'unit_wise':      
                      r2        mse       rmse      mape     smape        mae
component                                                                
passengers        0.866962  590.08557  24.291677  3.694311  3.743241  18.008002, 
'time_wise': {}}

print(signal.performance_models['unit_wise']['rmse'])

Output:

>>>               Prophet ExponentialSmoothing  AutoARIMA
component                                            
passengers     24.291677            40.306771   26.718103

Visualize predictions with the Visualization class:

Visualization(signal).show_predictions()

When models have been trained, you can compute predictions for future dates using the forecastmethod by passing it the name of a trained model and the horizon of prediction.

signal.forecast("Prophet", 100)
signal.forecast("AutoARIMA", 100)
signal.forecast("ExponentialSmoothing", 100)
Visualization(signal).show_forecast()

Aggregation of models

The method apply_aggregated_model aggregates the passed list of trained estimators according to their RMSE on train data. All passed models will be kept, so make sure to exclude models that are always less performant. The more a model performs compared to others, the greater its weight in agregation.

signal.apply_aggregated_model([AutoARIMA(), Prophet()])
signal.compute_scores(axis=1)
Visualization(signal).show_predictions()
signal.forecast("AggregatedModel", 100)

Author