AI in Procurement: Applying the ADF Test
Updated: Oct 20
When we analyze procurement data, we usually assume past patterns will guide future forecasts. But drift and instability in data can make that unreliable.
That's where the Augmented Dickey-Fuller (ADF) Test comes in. This statistical test helps determine whether a data series is stationary (meaning it exhibits a consistent statistical property over time) or non-stationary (meaning trends or seasonality may unpredictably influence the data).
Stationarity: What Does It Mean?
Stationarity implies that a time series fluctuates around a constant mean and variance over time, without exhibiting a long-term trend.
In contrast, non-stationary data exhibits trends, seasonality, or changing variances, making historical patterns unreliable for forecasting.
Why Does Stationarity Matter in Procurement?
Procurement relies heavily on forecasting—think lead times, supplier performance, and cost trends. If data drifts, traditional models like ARMA/ARIMA struggle, and your forecasts might be off.
As an example, consider supplier lead times. If historical data indicate a consistent increase in lead times due to deteriorating logistics conditions, relying on past averages for future planning may lead to stock shortages or urgent, high-cost orders.
Using the ADF test can help procurement teams assess whether a dataset is stable enough for forecasting or requires transformations (such as differencing) to remove trends and achieve stationarity.
How the ADF Test Works
The ADF test checks whether a time series has a unit root, which is a complex way of saying whether the data exhibits a trend that makes it non-stationary.
It does this by testing the null hypothesis that the time series has a unit root (i.e., is non-stationary).
If the p-value is below a certain threshold (typically 0.05), we reject the null hypothesis and conclude the series is stationary.
Don't be afraid; you won't need to remember all that. The outcome of the ADF test looks more or less like this:
"Interpreting the Results:
ADF Statistic: This value helps determine whether a unit root is present in the time series. A more negative value indicates more substantial evidence against the presence of a unit root, suggesting the series is stationary.
p-value: indicates how compatible the observed test result is with the null hypothesis of a unit root. A value below the chosen significance level supports rejecting that hypothesis. It is not the probability that the series is non-stationary.
Based on the provided data, the ADF test results are:
ADF Statistic: -1.95
p-value: 0.30
Given the p-value is above 0.05, we fail to reject the null hypothesis, indicating that the LME Copper prices for 2024 are likely non-stationary".
The ADF Test Formula
You can easily find that formula, and the purpose of this post isn't to copy and paste it for the sake of scientific appeal.
Why all the hassle?
In procurement, non-stationary data typically implies greater uncertainty and risk, requiring more agile and dynamic procurement strategies.
If the Data is Stationary
Reliable Forecasting – Standard statistical forecasting models (e.g., ARMA, ARIMA) work effectively.
Consistent Analysis – Historical trends remain relevant for future decision-making.
Stable Demand Patterns – Procurement strategies can rely on historical consumption data for inventory planning.
Price Predictability – Supplier pricing trends remain stable, enabling long-term contract negotiations.
If the Data is Non-Stationary
Advanced forecasting is required. Transformations (differencing, log transformations) or models like ARIMA must be applied.
Unstable Supplier Performance Trends – Requires deeper investigation into external factors affecting variability.
Changing Demand Patterns – Procurement needs flexible contracts and adaptive inventory management.
Market Volatility Impact – Procurement strategies must account for inflation, currency fluctuations, and supply chain disruptions.
Real-life example: Augmented Dickey-Fuller (ADF) Test on Brent Crude Oil Prices.
To assess the stationarity of Brent crude oil prices over the past 365 days, ChatGPT conducted the ADF test using daily closing prices from March 4, 2024, to March 4, 2025. The data was sourced from Yahoo Finance.
ADF Test Results:
Test Statistic: -1.892
p-value: 0.337
Critical Values:
1% level: -3.436
5% level: -2.864
10% level: -2.568
Interpretation:
The reported p-value of 0.337 exceeds the 5% significance level. The test statistic of −1.892 is also less negative than the 5% critical value of −2.864.
We therefore fail to reject the null hypothesis of a unit root. This result does not establish stationarity and should not be described as confirming it.
This historical example covers March 2024–March 2025. Reproducing the result requires the original price series, observation frequency, data-cleaning method, lag selection, and test specification.
AI is a chest of treasures.
An ADF result can inform the choice of forecasting method. It does not, by itself, determine whether an organization should hold inventory, hedge prices, or buy on the spot market. Those decisions also require analysis of operational requirements, storage costs, financial exposure and available contractual options.
Generally, I wouldn't have been doing the ADF test, as I don't know Python.
But now you can collect the data of your interest and feed it to any instance of AI you're using. You will receive the outcome immediately, either supporting or disapproving of your sourcing decision or category strategy. No programming skills or mathematical background required!
Isn't that a charm?
Comments