How an Outlier Flips a Leaderboard: The Mathematical Rift Between Z-Score and Min-Max Scaling

Inside a 2023 supply chain evaluation published in the Manufacturing Review, researchers observed a mathematical fragility that dictates modern leaderboards. When evaluating multiple suppliers across disparate metrics—cost in dollars, delivery speed in days, defect rates in percentages—a single supplier’s unusually high performance on one metric flipped the entire procurement ranking. The raw performance data had not changed, but the mathematical lens used to aggregate it had.

This is the fundamental problem of multi-criteria decision-making (MCDM) and machine learning preprocessing. You cannot add dollars to days. To combine them into a single score, the data must be normalized or standardized into a common, unitless scale. The two dominant methods for this translation are min-max scaling and z-score standardization, and choosing between them is not a neutral administrative step. It is an editorial decision that determines who wins.[1][2][5]

Min-max scaling is the more intuitive of the two. It takes the lowest value in a dataset and sets it to exactly 0, takes the highest value and sets it to exactly 1, and distributes everything else proportionally in between. The formula subtracts the minimum value from the observation, then divides by the range (the maximum minus the minimum).[2][4]

The vulnerability of this approach emerges the moment an extreme outlier enters the dataset. If 99 suppliers deliver a part in 10 to 12 days, but one supplier takes 150 days, the denominator of the min-max formula explodes. The 150-day supplier becomes the 1, and the 10-to-12 day cluster is violently compressed into a tiny band between 0.00 and 0.01. The variance that separates the normal suppliers is mathematically erased.[3][5]

As machine learning researcher Sebastian Raschka noted in his 2014 analysis of feature scaling, min-max scaling "bounds the data to a specific range" but leaves the dataset highly sensitive to these extreme values. When the algorithm downstream looks at the min-max scaled data, it sees one massive anomaly and a homogenous block of identical competitors, destroying the model's ability to differentiate among the pack.[4]

Z-score standardization takes a completely different approach to the denominator. Instead of looking at the absolute maximum and minimum, it centers the data around the mean (setting it to 0) and scales it by the standard deviation (setting the variance to 1). To calculate a z-score, you subtract the mean from the observation and divide by the standard deviation.[1][2]

Under z-score standardization, that same 150-day outlier does not compress the rest of the data. Because the standard deviation is calculated using all data points, the outlier receives a massive z-score—perhaps +6.5 or +7.0—while the 10-to-12 day suppliers maintain their relative distances from one another, clustering around -0.2 to +0.2.[3][5]

A 2026 curriculum update from upGrad emphasizes this distinction for machine learning models. Algorithms that compute distances between data points, like K-Nearest Neighbors (KNN) or Support Vector Machines (SVM), require standardized data to prevent a feature with a naturally large range (like a $500,000 salary) from overpowering a feature with a small range (like a 4.0 GPA).[1]

The DataCamp tutorial on the subject draws a hard line on when to deploy each. Standardization does not bound data to a specific range, which makes it much more robust to outliers, but it routinely produces negative numbers and lacks a defined maximum. Min-max scaling guarantees a strict 0-to-1 boundary, which is mandatory for algorithms like neural networks that expect bounded inputs, or image processing where pixel intensities must map cleanly to a 0-to-255 scale.[2]

The debate among practitioners on platforms like Stack Exchange often centers on the underlying distribution of the data. If the data roughly follows a Gaussian (bell-curve) distribution, z-score standardization is the mathematical default. If the distribution is entirely unknown or strictly non-Gaussian, min-max scaling provides a safer, albeit more brittle, containment strategy.[3]

The stakes of this mathematical choice extend far beyond server racks. University rankings, corporate ESG (Environmental, Social, and Governance) scores, and sovereign credit ratings all rely on aggregating disparate metrics. A shift from min-max to z-score can move a mid-tier university into the top 10 simply because its single massive endowment is allowed to pull its full weight, rather than being capped at 1.[5]

The architecture of a leaderboard is written in its denominator. When a final ranking is published, the most consequential variable is rarely the raw performance of the competitors. It is the mathematical boundary the architects chose to impose on the extremes.[5]