SIADS 523: Communicating Data Science Results · University of Michigan · December 2025
This project was completed as part of SIADS 523: Communicating Data Science Results at the University of Michigan. The course focuses on translating technical analysis into clear, audience-appropriate communication — written, visual, and oral — for stakeholders without a data science background.
The scenario centers on Acme Aroma, a manufacturing company experiencing a sustained rise in employee attrition — from 6% in 2012 to a peak of 19% in 2021, a 3x increase over the decade. The 2021–2022 period alone produced 1,355 departures and ₹29.8 million in recruitment and onboarding costs, the first acquisition expenses the company had incurred in over a decade. The project tasked me with building a predictive model to identify at-risk employees and delivering findings to the company's VP of Human Resources in a format she could act on.
Deliverables included a multi-slide technical report, a one-page executive summary, and a five-minute recorded presentation — all structured around the PAIRL communication framework (Problem, Approach, Insights, Recommendations, Limitations).
Skills demonstrated in this project
The executive summary condenses the full analysis into a single page for a non-technical stakeholder audience. It covers all five components of PAIRL — leading with the business problem and financial stakes, summarizing the analytical approach, presenting key model insights and odds ratios in plain language, and closing with the strategic recommendation and its projected cost savings.
The analysis drew from Acme's internal HRIS (Pegasus), containing 4,410 employee records across personal characteristics, professional history, and engagement survey results. A logistic regression model was selected for its dual utility — predicting which employees are at risk while also quantifying how strongly each factor drives that risk.
A critical data challenge was class imbalance: only 16% of records represented employees who left. The training data was balanced through oversampling to ensure the model learned meaningful patterns from both groups. Recall was prioritized as the primary evaluation metric, since missing a true departure is substantially more costly for the organization than a false alarm.
The final model included five statistically significant predictors. Employee engagement factors dominated — the top three were all within the company's direct control:
Five HR initiatives were evaluated by simulating each through the model. Limiting business travel — which directly improves work-life balance — produced the largest attrition reduction, lowering departure probability from 49.6% to 35.4% and retaining an estimated 563 employees annually. At 75% of annual salary as replacement cost, this translates to ₹318 million in avoided costs per year, outperforming the next-best option by roughly 20%.
The full technical report below covers each stage of the analysis in depth, including exploratory data analysis, model performance, the root association analysis, and the complete initiative comparison.
The findings were delivered as a five-minute recorded presentation built specifically for a non-technical executive audience. Rather than walking through the technical report, the presentation uses original data visualizations, a clear narrative arc, and plain language to lead the viewer from the problem through to the recommendation. Slides were designed separately from the technical report to support the spoken delivery rather than replicate it.