[Data Analysis]A product goes viral on social media, creating high demand. Distribution channels are facing supply chain issues because the testing…

This scenario is a classic example of data drift. Data drift occurs when the statistical properties of the production data, on which the model makes predictions, change or "drift" away from the data the model was originally trained on. The viral event caused a sudden, significant shift in the distribution of sales demand, making the historical training data no longer representative of the current reality. Consequently, the forecasting model's performance degrades because the patterns it learned are now obsolete.

A. Model bias: This refers to inherent, systematic errors in the model's predictions due to flawed assumptions, not a change in data distribution over time.

C. Incorrect sizing: This describes a potential consequence of the inaccurate forecast (e.g., insufficient inventory), not the underlying statistical problem with the model's data.

D. Skewing: This is a measure of a distribution's asymmetry. While the new data may be skewed, "data drift" is the specific term for the phenomenon of the data distribution changing over time.

---

1. Microsoft Azure Official Documentation. In the "What is data drift?" section for Azure Machine Learning

it states: "Data drift is one of the top two reasons for model accuracy degradation. It occurs in most machine learning applications... Data drift is the change in model input data that leads to model performance degradation." This directly aligns with the scenario where a change in demand patterns (input) causes the sales forecasting model to fail.

Source: Microsoft Corporation. (2023). Detect data drift (preview) on datasets. Azure Machine Learning Documentation. Section: "What is data drift?".

2. Google Cloud Official Documentation. The documentation for Vertex AI Model Monitoring defines data drift (also called feature skew) as occurring when the statistical properties of input features change over time. It distinguishes between training-serving skew and prediction drift. The scenario described is a form of prediction drift

where feature data distributions change between training and serving.

Source: Google Cloud. (2023). Introduction to Vertex AI Model Monitoring. Vertex AI Documentation. Section: "Data drift and concept drift".

3. Academic Publication. A foundational survey on the topic defines the problem: "In a changing environment

the data distribution may change over time

which is known as concept drift... This change may hurt the performance of a model trained on old data." The viral event represents a sudden change in the environment

causing the data distribution to shift.

Source: Gama

J.

Žliobaitė

I.

Bifet

A.

Pechenizkiy

M.

& Bouchachia

A. (2014). A survey on concept drift adaptation. ACM Computing Surveys (CSUR)

46(4)

1-37. Section 2: "Concept Drift". DOI: https://doi.org/10.1145/2523813

4. University Courseware. Stanford University's course on Machine Learning Systems Design explicitly covers this issue. Lecture notes describe data drift as a change in the input data distribution over time

which is a common reason for performance degradation in production ML systems.

Source: Stanford University. (2021). CS 329S: Machine Learning Systems Design

Lecture 4: Data. Course materials. Section: "Data Distribution Shifts".