Scalecast is a light-weight time-series forecasting procedure, wrapper, and results container built by and for applied Data Scientists using an ML framework. It offers a streamlined transforming, tuning, reverting, and reporting interface with many model classes, from basic ARIMA and linear models to boosted trees and recurrent neural nets. No matter which models you want to play with, the uniform interface makes it easy and fun to get results quickly.
All forecasts are validated out-of-sample, usually with a dynamic recursive approach (RNNs use a direct point estimate). You won't ever run into the situation where the estimator looks great on the test-set but can't generalize to real data. What you see is what you get, with no attempt to oversell results. If you run a model that's able to predict the COVID pandemic in your test-set, you probably have a one-step forecast happening under-the-hood, which is easy to fall into, especially for those not specialized in time series. You can't predict the unpredictable, and you won't see such things with scalecast.
from scalecast.Forecaster import Forecaster
from scalecast.Pipeline import Pipeline, Transformer, Reverter
from scalecast.auxmodels import mlp_stack
from scalecast import GridGenerator
import matplotlib.pyplot as plt
import pandas_datareader as pdr
models = (
'mlr',
'elasticnet',
'lightgbm',
'knn',
)
GridGenerator.get_example_grids()
df = pdr.get_data_fred(
'HOUSTNSA',
start='1959-01-01',
end='2022-08-01'
)
f = Forecaster(
y=df['HOUSTNSA'],
current_dates=df.index,
future_dates=24,
)
f.set_test_length(.2)
f.set_validation_length(24)
def forecaster(f,models):
""" add Xvars and forecast
"""
f.add_covid19_regressor()
f.auto_Xvar_select()
f.tune_test_forecast(
models,
dynamic_testing=24, # test-set metrics will be an average of rolling 24-step forecasts
cross_validate=True,
k = 3,
)
mlp_stack(f,models)
transformer = Transformer(
transformers = [
('DiffTransform',1),
('DiffTransform',12),
],
)
reverter = Reverter(
# list reverters in reverse order
reverters = [
('DiffRevert',12),
('DiffRevert',1),
],
base_transformer = transformer,
)
pipeline = Pipeline(
steps = [
('Transform',transformer),
('Forecast',forecaster),
('Revert',reverter),
],
)
f = pipeline.fit_predict(f,models=models)
f.reeval_cis() # expanding cis based on all model results
f.plot(ci=True,order_by='LevelTestSetMAPE')
plt.show()
results = f.export(
['model_summaries','lvl_fcsts']
)
The library provides the Forecaster
(for one series) and MVForecaster
(for multiple series) wrappers around the following estimators:
The Forecaster
object only can use:
The MVForecaster
object only can use:
Want more models? Open a feature request!
The library interfaces nicely with interactive notebook applications.
- Model Validation
- Probabilistic Forecasting and other confidence interval types (future: awesome conformal prediction)
- Model input analysis
- Anomaly detection
- Changepoint detection
- Series transformation/revert functions
- Only the base package is needed to get started:
pip install --upgrade scalecast
- Optional add-ons:
pip install darts
pip install prophet
pip install greykite
pip install shap
(SHAP feature importance)pip install kats
(changepoint detection)pip install pmdarima
(auto arima)pip install tqdm
(progress bar for notebook)pip install ipython
(widgets for notebook)pip install ipywidgets
(widgets for notebook)jupyter nbextension enable --py widgetsnbextension
(widgets for notebook)jupyter labextension install @jupyter-widgets/jupyterlab-manager
(widgets for Lab)
- Sklearn Univariate
- Sklearn Multivariate
- RNN
- ARIMA
- Theta
- VECM
- Other Notebooks