Business Intelligence & Analytics
A full analytics pipeline on a 50,000-row stress dataset that honestly reported 'no signal' instead of forcing a result.
The need
The brief was to find what actually drives stress and well-being at work, from a 50,000-row corporate dataset with 25 variables. The desire behind it: real, data-backed insight a business could act on.
The challenge
The data simply had no signal. The real test was integrity: resisting the strong temptation to torture the dataset until it produced a clean, publishable-looking story.
What I made
A full pipeline: visualisation, descriptive statistics, correlation analysis, regression, and machine-learning classification with decision trees and random forests, plus PCA and oversampling to handle class imbalance.
The outcome
We reported it straight: near-zero correlations, weak classification, and work-life balance not even statistically significant. The lesson, garbage in, garbage out, and the recommendation to collect better features rather than chase better models. The same honesty-first instinct that later became the AUC gate in Market Signal AI.
Key points
- Applied a full pipeline: visualisation, descriptive stats, regression, and ML classification
- Reported honest, counterintuitive 'no signal' findings instead of forcing results
- Diagnosed and tried to remediate class imbalance with oversampling
- Recommended collecting better features over optimising models on weak data