Skip to content

README.md

Latest commit

 

History

History
147 lines (100 loc) · 8.05 KB

README.md

File metadata and controls

147 lines (100 loc) · 8.05 KB

🌴 palm-churn-eda

2nd Place — BOLT BootCamp 2026

Alpha Boltlytics Consulting — EDA repository for diagnosing high-performer attrition at Palm Club Restaurants (Metro Vancouver, 7 branches).

Produced for BOLT BootCamp 2026 by Taksh Girdhar, Tanay Mahendru, Muskan Bhatia & Fiona Gandhi.

Context

Palm Club is a local restaurant chain with 7 locations across Metro Vancouver. Despite running quarterly performance reviews, the chain has a 75% overall employee turnover rate and an ~80% turnover rate specifically among high performers . This repository contains the full exploratory data analysis pipeline used to diagnose root causes and support data-driven recommendations to Palm Club management.

Key finding: This is a process problem, not a talent problem. The data to act already existed, it just wasn't being acted on.

Challenges Identified

Challenge 1 — Promotion delays for high performers. High performers wait a median of 7 months for their first promotion vs. 4 months for average performers, despite outscoring them. Over 75% of departing high performers cited better offers or lack of growth as their reason for leaving.

Challenge 2 — Inconsistent promotion criteria within and across branches. HP exit rates vary by 31 percentage points across branches (78.7% at UBC to 100% at Surrey and Downtown). Within the same branch, a lower-scoring employee (72.2) was promoted in 0.7 months while a higher scorer (92.7) waited 11.4 months.

Challenge 3 — No process to convert performance data into action. 33 high performers were never promoted despite having 3 or more performance reviews on file. 72% of part-time high performers were never promoted, compared to over 20% for full-time employees.

Recommendations

Recommendation 1 — Standardised Promotion Pipeline: Any employee in the top quartile (Bayesian-adjusted score) for 2 consecutive quarterly reviews is automatically flagged. Managers have a 30-day window to promote, defer with written reason, or initiate a 90-day development plan. Silence escalates to a regional manager.

Recommendation 2 — Manager Accountability & Training: HP retention rate added as a scored KPI in every manager's quarterly review. Managers trained to use a standardized promotion scorecard (performance score, attendance, peer feedback) replacing gut-feel decisions.

Target impact: Reduce HP turnover from 79.85% (214/268 exits) to 70% (188 exits), retaining 26 additional high performers across 7 branches.

Repository Structure

palm-churn-eda/
│
├── data/
│   ├── BOLT_Employees.csv
│   ├── BOLT_Applicants.csv
│   ├── BOLT_EmployeeChanges.csv
│   ├── BOLT_Performance.csv
│   └── BOLT_Branch.csv
│
├── Dataset_Profile_Reports/          # ydata-profiling HTML reports (one per table)
│
├── Outliers_Analysis/                # Outlier detection outputs and flagged records
│
├── palm-club-dashboard/              # Manager performance dashboard source code
│
├── load_data.ipynb                   # Data ingestion & initial inspection
├── Data_Profile_Generation.ipynb    # Automated dataset profiling via ydata-profiling
├── Outlier_Analysis.ipynb            # Outlier detection & flagging
├── Palm_Club_Emp_Turnover_EDA.ipynb # Main EDA notebook (9 guiding questions)
├── Checks.ipynb                      # SQL-based validation checks on employee_master
│
├── employee_master.csv               # Cleaned, merged master table (output of EDA pipeline)
├── index.html                        # Dashboard entry point
├── Dashboard README.md               # Dashboard setup and usage guide
└── README.md

Notebook Execution Order

The notebooks are designed to be run in sequence:

  1. load_data.ipynb — Ingests all five raw BOLT_*.csv tables and performs initial shape, dtype, and null inspection.
  2. Data_Profile_Generation.ipynb — Runs ydata-profiling on each table and saves interactive HTML reports to Dataset_Profile_Reports/.
  3. Outlier_Analysis.ipynb — Detects and flags statistical outliers; outputs saved to Outliers_Analysis/.
  4. Palm_Club_Emp_Turnover_EDA.ipynb — Main analysis notebook covering all 9 guiding questions (see table below).
  5. Checks.ipynb — Loads employee_master.csv into an in-memory SQLite database and runs SQL validation queries (e.g. education-level counts, churn cross-tabs).

employee_master.csv is the cleaned, merged output of the EDA pipeline. It can be loaded directly into any notebook to skip re-running the cleaning steps.

Analysis Coverage

Section Topic
3 Define "High Performer" (Bayesian-adjusted, top 25th percentile)
4 Churn rate: high performers vs. rest of workforce
5 Tenure-at-exit: when in their career do HPs leave?
6 Time-in-role before first promotion
7 Does promotion improve retention?
8 Branch-level HP churn differences
9 Wage tier vs. performance alignment
10 Working hours & exit risk
11 Full-time vs. part-time status & short-term intent

Key Definitions

High Performer — An employee whose Bayesian-adjusted average performance score places them in the top 25th percentile across the organization (threshold: Bayesian score > 82.97). Fired employees are excluded from churn analysis.

Churn — A record in BOLT_EmployeeChanges where NewRole indicates a voluntary exit or termination (fired employees excluded).

Reference Date — 2026-03-14. All tenure calculations for currently active employees use this as their end date.

Full-time vs. Part-time — Used as the primary segmentation variable in place of student vs. non-student status, as employment type proved more analytically predictive of churn behaviour.

Dashboard

The Manager Performance Dashboard provides real-time visibility into HP status, tenure stage, and promotion readiness across all 7 branches. Each dashboard feature maps directly to a challenge identified in the analysis:

Dashboard Feature Challenge Addressed
HP at-risk flag (90+ days in role without review) Promotion delays
FT vs. PT performance parity view Part-time employees overlooked
Standardised promotion scorecard No process to act on performance data
Manager KPI tracker (HP retention rate, quarterly) Lack of manager accountability

🔗 Live dashboard: https://takshg.github.io/palm-churn-eda/

See Dashboard README.md for local setup instructions.

Setup

pip install pandas numpy matplotlib seaborn ydata-profiling sqlalchemy
jupyter notebook

No external API keys or database connections required. All data is loaded from local CSV files in data/.

Notes & Pitfalls

  • Employees with no performance records will produce NaN for their Bayesian-adjusted score. Filter or handle separately before applying the high-performer flag.
  • BOLT_EmployeeChanges has multiple rows per employee. Use .groupby("EmployeeID").first() or .last() deliberately and document the choice.
  • Tenure for active employees is a lower bound capped at the reference date. Exclude active employees from tenure-at-exit charts; include them in retention analyses.
  • Fired employees are excluded from all churn and retention analyses per the team's stated assumptions.
  • Wage is categorical , not continuous. Encode before any correlation analysis.