validmind.ZivotAndrewsArch

ZivotAndrewsArch

@tags('time_series_data', 'stationarity', 'unit_root_test')

@tasks('regression')

defZivotAndrewsArch(dataset:validmind.vm_models.VMDataset):

Evaluates the order of integration and stationarity of time series data using the Zivot-Andrews unit root test.

Purpose

The Zivot-Andrews Arch metric is used to evaluate the order of integration for time series data in a machine learning model. It's designed to test for stationarity, a crucial aspect of time series analysis, where data points are independent of time. Stationarity means that the statistical properties such as mean, variance, and autocorrelation are constant over time.

Test Mechanism

The Zivot-Andrews unit root test is performed on each feature in the dataset using the ZivotAndrews function from the arch.unitroot module. This function returns several metrics for each feature, including the statistical value, p-value (probability value), the number of lags used, and the number of observations. The p-value is used to decide on the null hypothesis (the time series has a unit root and is non-stationary) based on a chosen level of significance.

Signs of High Risk

  • A high p-value suggests high risk, indicating insufficient evidence to reject the null hypothesis, implying that the time series has a unit root and is non-stationary.
  • Non-stationary time series data can lead to misleading statistics and unreliable machine learning models.

Strengths

  • Dynamically tests for stationarity against structural breaks in time series data, offering robust evaluation of stationarity in features.
  • Especially beneficial with financial, economic, or other time-series data where data observations lack a consistent pattern and structural breaks may occur.

Limitations

  • Assumes data is derived from a single-equation, autoregressive model, making it less appropriate for multivariate time series data or data not aligning with this model.
  • May not account for unexpected shocks or changes in the series trend, both of which can significantly impact data stationarity.