KPSS

Assesses the stationarity of time-series data in a machine learning model using the KPSS unit root test.

Purpose

The KPSS (Kwiatkowski-Phillips-Schmidt-Shin) unit root test is utilized to ensure the stationarity of data within a machine learning model. It specifically works on time-series data to establish the order of integration, which is essential for accurate forecasting. A fundamental requirement for any time series model is that the series should be stationary.

Test Mechanism

This test calculates the KPSS score for each feature in the dataset. The KPSS score includes a statistic, a p-value, a used lag, and critical values. The core principle behind the KPSS test is to evaluate the hypothesis that an observable time series is stationary around a deterministic trend. If the computed statistic exceeds the critical value, the null hypothesis (that the series is stationary) is rejected, indicating that the series is non-stationary.

Signs of High Risk

• High KPSS score, particularly if the calculated statistic is higher than the critical value.
• Rejection of the null hypothesis, indicating that the series is recognized as non-stationary, can severely affect the model’s forecasting capability.

Strengths

• Directly measures the stationarity of a series, fulfilling a key prerequisite for many time-series models.
• The underlying logic of the test is intuitive and simple, making it easy to understand and accessible for both developers and risk management teams.

Limitations

• Assumes the absence of a unit root in the series and doesn’t differentiate between series that are stationary and those border-lining stationarity.
• The test may have restricted power against certain alternatives.
• The reliability of the test is contingent on the number of lags selected, which introduces potential bias in the measurement.