sklearn-ts 0.0.6

Package for time series forecasting

Welcome to sklearn-ts

Testing time series forecasting models made easy :) This package leverages scikit-learn, simply tuning it where needed for time series specific purposes.

Main features include:

• Moving window time split
  • train-test split
  • CV on moving window time splits
• Model wrappers:
  • Neural networks

Other python packages in the time series domain:

• sktime
• sktime-dl
• darts

Installation

pip install sklearn-ts

Quickstart

Forecasting COVID-19 with Linear Regression

from sklearn_ts.datasets.covid import load_covid
from sklearn.linear_model import LinearRegression
from sklearn_ts.validator import check_model

dataset = load_covid()['dataset']
dataset['month'] = dataset['date'].dt.month

params = {'fit_intercept': [False, True]}
regressor = LinearRegression()

results = check_model(
    regressor, params, dataset,
    target='new_cases', features=['month'], categorical_features=[], user_transformers=[],
    h=14, n_splits=2, gap=14,
    plotting=True
)

Forecasting models

Model family	Model	Univariate
Benchmark	Naive	1
Exponential Smoothing	SES	1
Exponential Smoothing	Holt's linear	1
Exponential Smoothing	Holt-Winter	1
-	Prophet
Neural networks	ANN
Neural networks	LSTM
Neural networks	TCN

Documentation

Tutorial notebooks:

• neural networks

Development roadmap

• TCN przewaga
• Regularization
• XGBoost drawing
• FEATURES + SHAP
• x13
• prettier plot
• Handling many observations per date
• Constant window for forecasting
• For NN - chart of how it learned
• Logging
• Read the docs
• prod
• save picture optional
• PI Coverage
• Watermark
• OLS pi
• AIC / BIC penalizing coefficients / weights param vs hypreparams reg l1 l2, drop out, data augment, eartly stopping
• one step ahead forecast and again forecast etc
• pi for prophet - explaining how they are formulated
• tcn missing arrow
• tcn details
• t-test
• iterative one step ahead

JOURNAL

• daily but complicated -mae

• residuals normality as part of performance evaluation

• decide which measure to show

• those without features and pi still working

• czasem się nie przelicza - co wtedy? Zliczać błędne / 100?