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:
pip install sklearn-tsfrom 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
)
| 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 |
Tutorial notebooks:
- 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?