pmdarima AutoArima 0 mean (no intercept) with constant time series

AutoArima 0 mean (no intercept) with constant time series

Open BaklazhenkoNikita opened this issue 1 year ago • 3 comments

Describe the bug

When provided with a completely constant time series (0,0,0) ARMA is being created with no intercept. Hence, for all the constant time series we predict 0. Slight alteration of time series with epsilon on the other hand leads to (0,0,0) ARMA with mean intercept.

To Reproduce

import pandas as pd
from pmdarima import auto_arima
from IPython.display import display

#Testing the Arima predictions with constant time series (returns 0 predictions)
data = pd.Series([1]*20, index=pd.period_range(start='2024-01-01', periods=20, freq='D'))

auto_arima_model = auto_arima(data, error_action='ignore')

n_periods = 5 
forecast = auto_arima_model.predict(n_periods=n_periods)

display(auto_arima_model.summary())
print("Forecast for constant time series:", forecast)

#Testing the Arima predictions with constant time series and one altered value (returns Mean as predictions)
data_non_constant = pd.Series([1]*19 + [1.1], index=pd.period_range(start='2024-01-01', periods=20, freq='D'))

auto_arima_model_non_constant = auto_arima(data_non_constant, error_action='ignore')

n_periods = 5 
forecast = auto_arima_model_non_constant.predict(n_periods=n_periods)

display(auto_arima_model_non_constant.summary())
print("Forecast for almost constant time series:", forecast)

Versions

Linux-5.10.16.3-microsoft-standard-WSL2-x86_64-with-glibc2.31
Python 3.9.11 (main, Mar 17 2022, 06:00:02) 
[GCC 10.2.1 20210110]
pmdarima 1.8.4
NumPy 1.24.4
SciPy 1.11.4
Scikit-Learn 1.2.2
Statsmodels 0.14.1