Analytics is not about charts. It is about changing what a leader does next. This page walks through one complete analytical story — from the business question a leader asked, through exploration, pattern identification, driver analysis, and the insight that changed how the organization acted.
Integrity & Confidentiality: All data, percentages, region names, and identifiers shown are entirely synthetic. No real employees, locations, or proprietary data from any organization are represented. Analytical frameworks and methodologies reflect large-scale workforce analytics and transformation work.
"Why are some locations underperforming? We have the same product, the same training. What is different?"
The question
A leader needed an answer
Not a dashboard. A specific answer to a specific business problem — with enough evidence to act on it confidently.
The method
Data before conclusions
Explore first. Find patterns before building models. Let the data reveal what is worth investigating.
The outcome
The intervention changed
Leaders stopped watching turnover. They started watching absence — 90 days earlier in the cycle.
Every analysis follows the same six-step discipline. The question determines the analysis — not the other way around. The most common mistake is jumping from step two to step six. Everything in between is where the insight lives.
Decision Science Workflow
A clearer, portfolio-ready version of the analytical flow — from business question to executive decision.
1
Frame the Business Question
Translate an ambiguous request into a well-defined problem. Clarify priorities, decision points, and what action the analysis must influence.
2
Identify the Available Data
Map internal, external, survey, operational, and web-based datasets. Confirm what is reliable, missing, or unavailable.
3
Clean and Prepare the Data
Examine fields at a detailed level, treat missing values, standardize definitions, and create analysis-ready data.
4
Explore Patterns and Features
Segment the data, compare groups, test relationships, and validate observations with business stakeholders.
5
Model, Evaluate, and Refine
Build statistical or machine-learning models only when they help answer the business question. Evaluate, refine, and pressure-test results.
6
Communicate the Decision Story
Visualize findings, explain trade-offs, and translate the insight into a clear recommendation leaders can act on.
Cleaned up visual: from business question to executive decision
This replaces the blurry image with responsive HTML/CSS, so the text stays sharp and fits the page on desktop and mobile.
Frame before collecting
The question determines which data matters. Collecting data without a question produces noise, not insight.
Explore before modelling
A scatter matrix and correlation heatmap before any model. What the data is actually saying — before assumptions take over.
Story before slides
The visual is the last step. Designed around the decision a leader needs to make — not around what looks impressive.
Chapter 02
Data exploration — before any model
Before building anything, the data needs to speak. A scatter matrix reveals distributions and pairwise relationships. A correlation heatmap shows which variables move together. A dendrogram identifies natural clusters. This is Python work — and it determines whether everything that follows will be meaningful.
Synthetic data only. Variable names, values, and patterns shown are generated for demonstration. No real organizational data represented.
LOA rate, tenure, and internal hire rate cluster together. Engagement moves with them. This is not coincidence — it is a system.
Diagonal: distribution of each variable. Off-diagonal: pairwise scatter. Built in Python before any model is considered.
What relationships exist between workforce variables?
Years in role, years in organization, age, and shift count. The distributions and scatter patterns determine which variables are worth including in any model.
Triangular heatmap showing similarity scores. Green intensity = stronger correlation. Three clusters emerged.
Which variables move together — and how do they cluster?
Left: triangular correlation matrix. Top right: dendrogram from hierarchical clustering. Bottom right: actual vs predicted time series across three workforce metrics.
What we found
Tenure and IHR have the strongest positive co-movement. LOA and turnover cluster together. Engagement is downstream — a lagging indicator, not a driver.
Why it changed the approach
Had we skipped exploration and modelled directly, we would have included engagement as a feature. The heatmap showed it moves after LOA — not before.
The Python proof
Seaborn pairplot, scipy hierarchical clustering, matplotlib time series overlay. Built in Google Colab before a single Power BI visual was designed.
Chapter 03
Where is the problem — and who is most affected?
Two questions answered simultaneously. The geo-risk map shows where risk concentrates across the retail network. The engagement heatmap shows who is most affected — by role level and by tenure band. Together they make the problem undeniable.
Synthetic data only. All coordinates, risk scores, percentages, and identifiers are generated for demonstration. No real locations or employee data represented.
Risk concentrates in specific geographies and specific populations — manager-level employees in their first three years are the critical cohort.
Workforce Risk Landscape
Each dot represents a synthetic location. Colour indicates risk tier. No identifiable location-level data is shown.
Low riskModerate riskHigh riskCritical risk
Engagement by role level — seven pillars
Single blue ramp. Darker = stronger. Manager level is the consistent weak point.
Engagement by tenure band — the same seven pillars
Scores strong in first three months. Drop sharply in year one through three. Partially recover after five years. Year one is the critical window.
Chapter 04
Why it happens — three pillars, one system
The heatmap showed where and who. The driver analysis showed why. Performance is shaped by three interconnected pillars — each requiring a different intervention. No single factor dominates everywhere.
Synthetic data only. All percentages shown are generated for demonstration.
Roughly 1 in 5 locations underperform. Each is primarily driven by one of three pillars — but the mix varies by region. One intervention strategy does not fit all.
Pillar 1
Employee Experience
Engagement, belonging, development, and management quality. Widespread opportunity — not the primary driver in most locations.
~32%
of underperforming locations
Pillar 2
Workforce Stability
LOA rate, turnover, and staffing availability. Driven by both absence pressure and elevated exits occurring simultaneously.
~35%
of underperforming locations
Pillar 3
Leadership Continuity
Experience in role, organizational tenure, in-location consistency. The upstream driver that shapes everything below it.
~33%
of underperforming locations
Driver distribution by region — stacked view (synthetic)
Power BI — workforce intelligence at enterprise scale
Four dashboards built at Walmart Canada. Each answers one specific question. Top left: store performance with RAG scoring. Top right: workforce composition trends. Bottom left: ER incident geo-mapping across Canada. Bottom right: compliance and investigation tracking.
From pattern detection to executive dashboard — the full pipeline in production
Built at Walmart Canada. The geo-risk map (bottom left) directly applies the same geographic risk plotting methodology shown in Chapter 03 — real enterprise output of the same analytical approach.
Chapter 05
The full picture
Three pillars. One causal chain. And when that chain breaks — a predictable cycle that repeats until someone intervenes at the right point. This is what the analysis produced: a model simple enough to act on.
Synthetic data only. All metrics and percentages are generated for demonstration.
Internal hiring rebounds. Turnover begins to fall.
Signal 6 · Reset
↑ ENG
Engagement lifts. Cycle closes — until LOA rises again.
Watch LOA — not turnover. Intervene 90 days earlier. That is the insight that changed the strategy.
52-week rolling trend — the cycle made visible (synthetic)
LOA spikes first. IHR drops next. Turnover follows. Once you see it you cannot make the same decisions.
LOA RateTurnoverInternal Hire RateEngagement
What changed
The organization shifted from monitoring turnover as the primary attrition signal to monitoring LOA rate. Intervention triggers were set at LOA thresholds — before turnover had a chance to spike.
The lesson
The most valuable insight is rarely in the metric everyone is already watching. It is in the upstream signal nobody thought to track. Finding that signal is what decision science is actually for.
