How to Pinpoint the Real Drivers Behind Metric Fluctuations: Methods & Case Studies

What is Metric Attribution

Metric attribution is the process of locating the core factors that cause a metric’s fluctuation, whether the change is a rise, drop, or sudden spike. It follows three main steps: clarify the problem, analyze and locate the cause, and finally solve the issue.

1. Clarify the Problem

Determine whether the observed change is significant and whether it falls within the scope of metric attribution. This may involve questioning why a metric is declining, whether the decline signals an unsustainable pattern, or what positive changes are driving an increase.

2. Analyze and Locate the Problem

Select an appropriate analytical method to quantify the contribution of each factor. Methods include deterministic judgment, probabilistic judgment, and speculative judgment, which can be combined as needed.

3. Solve the Problem

Translate the analytical findings into concrete business actions, such as adjusting product features or marketing strategies based on the identified drivers.

Basic Methods of Metric Attribution

1. Metric Judgment

First verify that the metric change is real before proceeding with attribution.

2. Attribution Methods

Three categories of analysis:

Deterministic judgment – e.g., metric decomposition to allocate impact to specific components.

Probabilistic judgment – use modeling (machine learning, SHAP values) to estimate each factor’s contribution.

Speculative judgment – generate hypotheses about possible causes and later validate them with deterministic or probabilistic methods.

3. Deterministic Judgment

Metric decomposition can be performed using addition, subtraction, multiplication, or division:

Addition : Split a metric (e.g., revenue) into contributions from different channels.

Subtraction : Similar to addition but with opposite signs.

Multiplication : Decompose a conversion funnel, assigning a weight to each stage; includes replacement method and LMD product‑factor decomposition.

Division : Used for ratio‑type metrics; can be handled via product‑factor decomposition or a dual‑factor method (structural + fluctuation contributions).

These methods can be combined for more comprehensive analysis.

Case Studies

Case 1 – Deterministic Judgment via Metric Decomposition

A monitoring alert showed a 3.13% drop in a conversion rate with a -27.8% swing. Using multiplication‑based decomposition, the analysis identified stages D and B as the main negative contributors, while stage C had a positive effect. The study emphasized the need for meaningful, business‑relevant metrics and cautioned against over‑splitting into too many sub‑metrics.

Further drilling down by dimension (e.g., gender or page section) helped pinpoint specific sub‑scenarios, guiding targeted strategy adjustments.

Case 2 – Probabilistic Judgment with Machine Learning & SHAP

An observed 6.7% increase in user activity prompted a probabilistic analysis. A predictive model was built, and SHAP values were computed to rank feature importance. Feature A showed the strongest positive impact on the activity metric. Visualizations confirmed a linear positive relationship between feature A’s value and the target metric.

The insight led to business hypotheses (e.g., content excitement) and subsequent strategic recommendations.