Skip to content

A python package for time series forecasting with scikit-learn estimators.

License

Notifications You must be signed in to change notification settings

cerlymarco/tspiral

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

23 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

tspiral

A python package for time series forecasting with scikit-learn estimators.

tspiral is not a library that works as a wrapper for other tools and methods for time series forecasting. tspiral directly provides scikit-learn estimators for time series forecasting. It leverages the benefit of using scikit-learn syntax and components to easily access the open source ecosystem built on top of the scikit-learn community. It easily maps a complex time series forecasting problems into a tabular supervised regression task, solving it with a standard approach.

Overview

tspiral provides 4 optimized forecasting techniques:

  • Recursive Forecasting

Lagged target features are combined with exogenous regressors (if provided) and lagged exogenous features (if specified). A scikit-learn compatible regressor is fitted on the whole merged data. The fitted estimator is called iteratively to predict multiple steps ahead.

recursive-standard

Which in a compact way we can summarize in:

recursive-compact

  • Direct Forecasting

A scikit-learn compatible regressor is fitted on the lagged data for each time step to forecast.

direct

Which in a compact way we can summarize in:

direct-compact

It's also possible to mix recursive and direct forecasting by predicting directly some future horizons while using recursive on the remaining.

directmix-compact

  • Stacking Forecasting

Multiple recursive time series forecasters are fitted and combined on the final portion of the training data with a meta-learner.

stacked

  • Rectified Forecasting

Multiple direct time series forecasters are fitted and combined on the final portion of the training data with a meta-learner.

rectify

GLOBAL and MULTIVARIATE time series forecasting are natively supported for all the forecasting methods available. For GLOBAL forecasting, use the groups parameter to specify the column of the input data that contains the group identifiers. For MULTIVARIATE forecasting, pass a target with multiple columns when calling fit.

Installation

pip install --upgrade tspiral

The module depends only on NumPy, Pandas, and Scikit-Learn (>=0.24.2). Python 3.6 or above is supported.

Media

Usage

  • Recursive Forecasting
import numpy as np
from sklearn.linear_model import Ridge
from tspiral.forecasting import ForecastingCascade
timesteps = 400
e = np.random.normal(0,1, (timesteps,))
y = np.concatenate([
    2*np.sin(np.arange(timesteps)*(2*np.pi/24))+e,
    2*np.cos(np.arange(timesteps)*(2*np.pi/24))+e,
])
X = [[0]]*timesteps+[[1]]*timesteps 
model = ForecastingCascade(
    Ridge(),
    lags=range(1,24+1),
    groups=[0],
).fit(X, y)
forecasts = model.predict([[0]]*80+[[1]]*80)
  • Direct Forecasting
import numpy as np
from sklearn.linear_model import Ridge
from tspiral.forecasting import ForecastingChain
timesteps = 400
e = np.random.normal(0,1, (timesteps,))
y = np.concatenate([
    2*np.sin(np.arange(timesteps)*(2*np.pi/24))+e,
    2*np.cos(np.arange(timesteps)*(2*np.pi/24))+e,
])
X = [[0]]*timesteps+[[1]]*timesteps 
model = ForecastingChain(
    Ridge(),
    n_estimators=24,
    lags=range(1,24+1),
    groups=[0],
).fit(X, y)
forecasts = model.predict([[0]]*80+[[1]]*80)
  • Stacking Forecasting
import numpy as np
from sklearn.linear_model import Ridge
from sklearn.tree import DecisionTreeRegressor
from tspiral.forecasting import ForecastingStacked
timesteps = 400
e = np.random.normal(0,1, (timesteps,))
y = np.concatenate([
    2*np.sin(np.arange(timesteps)*(2*np.pi/24))+e,
    2*np.cos(np.arange(timesteps)*(2*np.pi/24))+e,
])
X = [[0]]*timesteps+[[1]]*timesteps 
model = ForecastingStacked(
    [Ridge(), DecisionTreeRegressor()],
    test_size=24*3,
    lags=range(1,24+1),
    groups=[0],
).fit(X, y)
forecasts = model.predict([[0]]*80+[[1]]*80)
  • Rectified Forecasting
import numpy as np
from sklearn.linear_model import Ridge
from sklearn.tree import DecisionTreeRegressor
from tspiral.forecasting import ForecastingRectified
timesteps = 400
e = np.random.normal(0,1, (timesteps,))
y = np.concatenate([
    2*np.sin(np.arange(timesteps)*(2*np.pi/24))+e,
    2*np.cos(np.arange(timesteps)*(2*np.pi/24))+e,
])
X = [[0]]*timesteps+[[1]]*timesteps  
model = ForecastingRectified(
    [Ridge(), DecisionTreeRegressor()],
    n_estimators=24*3,
    test_size=24*3,
    lags=range(1,24+1),
    groups=[0],
).fit(X, y)
forecasts = model.predict([[0]]*80+[[1]]*80)

More examples in the notebooks folder.