EU AI Act Compliance — Read our original regulation brief on how the EU AI Act aims to balance innovation with safety and accountability, setting standards for responsible AI use

ADF

Assesses the stationarity of a time series dataset using the Augmented Dickey-Fuller (ADF) test.

Purpose

The Augmented Dickey-Fuller (ADF) test metric is used to determine the order of integration, i.e., the stationarity of a given time series dataset. The stationary property of data is pivotal in many machine learning models as it impacts the reliability and effectiveness of predictions and forecasts.

Test Mechanism

The ADF test is executed using the adfuller function from the statsmodels library on each feature of the dataset. Multiple outputs are generated for each run, including the ADF test statistic and p-value, count of lags used, the number of observations considered in the test, critical values at various confidence levels, and the information criterion. These results are stored for each feature for subsequent analysis.

Signs of High Risk

  • An inflated ADF statistic and high p-value (generally above 0.05) indicate a high risk to the model's performance due to the presence of a unit root indicating non-stationarity.
  • Non-stationarity might result in untrustworthy or insufficient forecasts.

Strengths

  • The ADF test is robust to sophisticated correlations within the data, making it suitable for settings where data displays complex stochastic behavior.
  • It provides explicit outputs like test statistics, critical values, and information criterion, enhancing understanding and transparency in the model validation process.

Limitations

  • The ADF test might demonstrate low statistical power, making it challenging to differentiate between a unit root and near-unit-root processes, potentially causing false negatives.
  • It assumes the data follows an autoregressive process, which might not always be the case.
  • The test struggles with time series data that have structural breaks.