VarGroup — Customer Reactivation Propensity Model

A machine learning project that ranks dormant B2B clients by their probability of returning, turning a 2-million-row transaction log into a campaign-ready priority list for marketing.

---

Business problem

VarGroup's marketing team needed a way to decide which inactive customers to re-engage first. With tens of thousands of dormant clients and a finite campaign budget, random outreach wastes money on clients who will never come back. The goal of this project was to build a reactivation propensity score that ranks every silent client by how likely they are to make a positive-NET purchase if contacted — so the sales team can focus on the top of the list.

---

TL;DR results

Metric	Value
Inactive clients scored	13,244
Reactivations observed in target window	875 (6.61%)
Best model	Extra Trees Classifier
AUC-ROC	0.70
PR-AUC (Average Precision)	0.22
F1-optimal threshold	0.55
Lift over random targeting	3.4×

For every 100 clients contacted using the model, the campaign reaches 3.4× more reactivatable customers than a random outreach of the same size.

---

Methodology

1. Time-split target definition (no leakage)

• Observation cutoff: 2019-12-31
• Feature window: all transactions strictly before the cutoff
• Target window: January–June 2020 (6 months after cutoff)
• Target = 1 if the client has at least one positive-NET transaction in the target window; returns and zero-value lines do not count
• Population filter: only clients with no purchases in the 24 months before the cutoff are included — i.e. the model learns the transition from inactive to active, not the next purchase of an already-engaged customer

2. Feature engineering

Transactions are aggregated from 2M rows to one row per client (an Analytical Base Table), capturing:

• RFM core — recency, frequency, monetary sum/mean/max/std
• Inter-purchase gaps — mean, max, and standard deviation of days between purchases
• Product mix — unique items, unique families, entropy of purchase categories
• Spend windows — last-12-month and last-36-month spend
• Tool vs. non-tool split — spend share and transaction share on core tool products
• Spend momentum — monthly spend trend over the client's history
• Ratios — average ticket, spend per active month, tool share of transactions
• Firmographics — region, trade sector, employee count, economic potential, potential gap

3. Class imbalance

Reactivators are ~6.6% of the population. Rather than using SMOTE (which can inject noise into tabular data), random undersampling of the majority class was used. Ratios 1:5, 1:10 and 1:20 were tested; 1:10 gave the best Precision–Recall trade-off.

4. Modeling

Three models trained on the balanced set and evaluated on the original imbalanced test set (so metrics reflect real deployment):

• Extra Trees Classifier ← selected
• XGBoost
• Logistic Regression

Selection criterion: PR-AUC, the most informative metric under heavy class imbalance.

5. Threshold tuning

A threshold sweep from 0.10 to 0.90 was used to map the Precision/Recall trade-off. The F1-maximising threshold of 0.55 is the default operating point, but it is deliberately exposed as a business lever: lower it to cast a wider net, raise it to prioritise conversion rate over reach.

6. Feature importance & leakage audit

• Both native (impurity-based) and permutation importance were computed — permutation is model-agnostic and more robust against high-cardinality bias
• A dedicated leakage audit retrained the model without registry snapshot fields (columns ending in _first) and compared AUC delta, to verify that results are not driven by information that leaked from the target window into the static client registry

7. Scoring & delivery

The final model is applied to the full inactive population to produce reactivation_scores.csv — every client ranked by predicted reactivation probability. The subset above threshold is written to campaign_targets.csv, ready for the CRM.

---

Key business insights

• High-spend clients have the highest reactivation propensity. The "whales" that have gone quiet are disproportionately worth chasing.
• Tool-focused buyers (60–80% tool spend share) are the best campaign targets. Pure generalists and pure specialists are harder to bring back.
• Recency alone does not predict reactivation. Behavioural patterns — product diversity, spend momentum, inter-purchase variability — matter far more than "how long since they last bought".
• The top-ranked features by permutation importance are TRADE SECTOR, REGION, and MAX_GAP_DAYS — a mix of firmographics and behavioural variability.

---

Project structure

VarGroup-Reactivation-Model/
├── README.md
├── requirements.txt
├── .gitignore
├── notebooks/
│   └── project_final.ipynb       # Full end-to-end analysis
├── outputs/
│   ├── reactivation_scores.csv   # All 13,244 inactive clients, ranked
│   └── campaign_targets.csv      # Subset above the F1-optimal threshold
└── data/
    └── README.md                 # Data is confidential and not shipped

---

Data

The raw transactional data (TOOL_CLIENT.csv, TOOL_SALES.csv) is not included in this repository — it contains confidential client information. The notebook assumes both files are placed in data/ before running.

The shipped outputs/ files contain anonymised client IDs and model scores only, no transactional detail.

---

How to run

# 1. Clone
git clone https://github.com/<your-username>/VarGroup-Reactivation-Model.git
cd VarGroup-Reactivation-Model

# 2. Create a virtual environment (recommended)
python -m venv .venv
source .venv/bin/activate          # macOS / Linux
# .venv\Scripts\activate           # Windows

# 3. Install dependencies
pip install -r requirements.txt

# 4. Place the raw data in data/ (not shipped — see Data section)
#    data/TOOL_CLIENT.csv
#    data/TOOL_SALES.csv

# 5. Open the notebook
jupyter lab notebooks/project_final.ipynb

---

Tech stack

• Python 3.10+
• pandas, numpy — data wrangling
• scikit-learn — Extra Trees, Logistic Regression, pipelines, metrics, permutation importance
• xgboost — gradient-boosted trees
• category_encoders — target encoding for high-cardinality categoricals
• matplotlib, seaborn — EDA and result visualisation

---

Future work

• Calibration — Platt scaling or isotonic regression to make scores directly interpretable as probabilities
• Uplift modeling — move from "who will reactivate?" to "who will reactivate because we contacted them?", which is the real marketing question
• Time-aware cross-validation — rolling-origin CV across multiple observation dates for a more robust generalisation estimate
• Survival analysis framing — model time-to-reactivation rather than a binary 6-month window
• Cost-sensitive thresholding — replace F1 with a business-loss function that weighs campaign cost against lifetime value

---

License

This project is released under the MIT License — see LICENSE for details.

---

Built as an end-to-end case study in customer analytics: from raw transactions to a deployed marketing priority list.