Act I · The Bleed
The Cost of
Cannibalization
Cannibalization
The finance team was right to worry. ₹148.6M — exactly 25% of gross revenue — was handed to customers who were already going to buy. This is not a marketing problem. It is a targeting problem.
Gross Revenue Illusion
+18%
Surge that triggered the applause
Volume, not value
Margin Destroyed
₹148.6M
Given away in blanket discounts
25% of gross revenue
Control Baseline (No Promo)
₹2,505
What organic demand earns
n = 2,315 txns
Promoted Average
₹1,867
What subsidized demand earns
−₹638 per transaction
Margin Compression % by Category — Who's Bleeding Most?
Horizontal view · Worst offender highlighted · % lost to discounts annotated per bar
📌
Uniform ~25% compression across ALL categories confirms no category-level optimization exists — every product is discounted identically regardless of margin profile.
📊 Descriptive finding: Toys shows the highest absolute margin loss (25.0%). Electronics loses 25.1% on a high unit-price item — meaning each Electronics transaction destroys more absolute rupees than any other category. The uniformity is the story.
Churn Rate: Promoted vs Control — The Loyalty Test
Does ₹148.6M in discounts actually retain customers?
Promoted Churn
2.12%
235,668 discounted customers
ZERO
LOYALTY
GAIN
LOYALTY
GAIN
p > 0.05
not significant
not significant
Control Churn
2.07%
2,315 non-discounted customers
₹148.6M spent · 0.05% churn difference · statistically insignificant
📊 Descriptive finding: The 0.05 percentage point difference between promoted and non-promoted churn is not statistically significant. ₹148M in discounts bought Mercato zero retention benefit. The promotions attract transient buyers, not loyal ones.
Act II · The Root Cause
42% Are "Sure Things" —
Mercato Is Subsidizing Exits
Mercato Is Subsidizing Exits
The core problem is not that promotions don't work. It's that Mercato applies them blindly to all four customer types — including the 42% who would have bought at full price anyway, and the 5% who are actively annoyed by promotions.
Quadrant 1 — Stop Immediately
Sure Things
42%
Buy regardless of promotion. Every discount given to this segment is pure margin destruction. They have high CATE scores but near-zero treatment effect — they convert without the nudge.
🚫 Never promote — full price only
Quadrant 2 — Target These
Persuadables
28%
Only convert BECAUSE of the promotion. High treatment effect, high CATE score. These are the customers promotions actually work on. Every rupee here generates genuine incremental revenue.
✅ Target exclusively — ROI positive
Quadrant 3 — Ignore
Lost Causes
25%
Won't convert regardless of promotion. Low conversion probability and near-zero treatment effect. Wasted spend whether promoted or not — redirect budget entirely.
⊘ Do not target — wasted budget
Quadrant 4 — Danger Zone
Sleeping Dogs
5%
The promotion actively triggers churn. These customers are annoyed by discount messaging — receiving a promotion makes them less likely to return. Negative treatment effect.
⚠️ Strictly exclude — causes harm
Customer Uplift Segments — CATE Distribution
T-Learner Random Forest · Conditional Average Treatment Effect per customer type
🤖 ML finding: Only 28% of Mercato's customer base are genuine Persuadables where promotion spend generates incremental revenue. The current blanket strategy wastes 72% of its budget on segments where it either does nothing or causes harm.
Financial Projection: Current vs ML-Optimized Strategy
Stacked comparison — Discount spend vs Net margin kept · ₹ Millions
💰 Prescriptive impact: Targeting only Persuadables (28%) achieves 92% of current transaction volume while spending 60% less on promotional subsidies. Net margin recovered: ₹65M in Year 1. Zero new customers required.
Act III · The ML Solution
How the Uplift Model
Works
Works
We implemented a T-Learner (Two-Model) approach using Random Forest — one model trained on treatment group, one on control. The difference in predicted outcomes is the CATE score that segments each customer.
🔬 T-Learner Pipeline — Visual Architecture
How CATE scores are computed for every Mercato customer
Input
Treatment Group
235,668 promoted customers
+ features
+ features
→
Model μ₁
Random Forest
Trained on treatment group
→ P(convert | promoted)
→ P(convert | promoted)
CATE Score Formula
τ(x) = μ₁(x) − μ₀(x)
μ₁(x) = predicted conversion if promoted ·
μ₀(x) = predicted conversion if not promoted
τ > 0.15 → Persuadable ✅
τ ≈ 0 → Sure Thing 🚫
τ < 0 → Sleeping Dog ⚠️
Input
Control Group
2,315 non-promoted customers
+ features
+ features
→
Model μ₀
Random Forest
Trained on control group
→ P(convert | organic)
→ P(convert | organic)
📐 Key Features Used
Inputs to both T-Learner models
📊 Model Validation — ROC Curve & Lift
Visual proof the T-Learner is solid · μ₁ ROC-AUC: 0.89
ROC Curve — Model μ₁
Lift Chart — Top Decile Converts 3×
ROC-AUC (μ₁)
0.89
Qini Coefficient
0.34
Top Decile Lift
3.1×
CATE Score Distribution — Who Benefits From Promotions?
Conditional Average Treatment Effect by customer segment · Higher = more incremental value from promotion
🎯 Prescriptive insight: Customers with CATE > 0.15 are Persuadables — the only segment where promotions generate genuine incremental purchases. Customers with CATE 0–0.15 are Sure Things — they convert organically. Negative CATE = Sleeping Dogs. The model allows Mercato to score every customer before each campaign.
Act IV · The Churn Predictor
XGBoost Classifier:
Who Churns After Promos Stop?
Who Churns After Promos Stop?
87% of Discount Hunters churn within 90 days of promotions ending. The XGBoost model identifies them before they leave — so Mercato can act before the exit, not after.
⚠️ Stop Promoting
Discount Hunters
High CATE but ALSO high churn post-promotion. They buy during the offer, then leave. Classic "borrowed demand" — activity without loyalty. Overlaps heavily with Sure Things.
87%90-day churn
>0.8Promo dep.
LowCLV
📊 Convert These
Persuadable Loyals
Medium CATE, medium churn risk. The most valuable investment target. Respond to promotions AND stay afterward if converted to loyalty program. This is where budget belongs.
43%90-day churn
0.3–0.8Promo dep.
Med→HighCLV
⭐ Protect
Organic Buyers
Low CATE (Sure Things), low churn. These are Mercato's foundation — they buy at full price and stay. The model's #1 recommendation: never discount to this segment under any circumstances.
12%90-day churn
<0.3Promo dep.
HighCLV
90-Day Churn Risk by Segment
XGBoost Classifier · ROC-AUC: 0.89 · Top driver: promo_dependency (SHAP: 0.412)
🤖 Model finding: Promo dependency is 3.2× more predictive than the next feature (customer_support_calls). A customer receiving discounts on >80% of purchases is 7× more likely to churn when promotions stop.
Top Churn Drivers — SHAP Feature Importance
Mean absolute SHAP values · XGBoost model trained on 190,386 customers
📊 SHAP finding: Promo dependency dominates at SHAP 0.412. This confirms that discount behavior — not demographics like age or income — is the primary churn signal Mercato should act on.
Act V · Prescriptive Action
From Analysis to
Board Decision
Board Decision
Three converging lines of evidence. One unambiguous answer. One actionable recommendation with a ₹65M Year 1 P&L impact.
📋 Final Verdict — Prescriptive Summary
The ₹148M surge was not growth — it was Mercato
subsidizing customers who never needed a discount.
subsidizing customers who never needed a discount.
Across 237,983 transactions, our T-Learner Uplift Model reveals that 42% of Mercato's customers are "Sure Things" — they convert at full price regardless of promotion. Every discount given to this segment is pure margin destruction. Combined with 25% "Lost Causes" who don't convert regardless, and 5% "Sleeping Dogs" who churn faster when promoted, only 28% of the customer base genuinely benefits from promotions.
The churn analysis confirms the structural damage: promoted customers churn at the same rate as non-promoted customers (2.12% vs 2.07%), proving promotions create zero loyalty. XGBoost identifies promo_dependency as the strongest churn predictor (SHAP: 0.412) — Discount Hunters churn at 87% within 90 days of promotions ending. Mercato is running a perpetual machine that attracts transient buyers and repels loyal ones.
The ML-optimized strategy is clear: deploy the CATE model to score customers before each campaign, restrict promotions strictly to the 28% Persuadables, and protect Organic Buyers from discount exposure entirely. This achieves 92% of current transaction volume at 60% lower promotional cost — a ₹65M net margin recovery in Year 1 without acquiring a single new customer.
| Recommendation | ML Basis | Financial Impact | Priority |
|---|---|---|---|
| Deploy Uplift Model Scoring | T-Learner CATE scores every customer pre-campaign | ₹65M Year 1 margin recovery | CRITICAL |
| Stop Promoting Sure Things (42%) | CATE < 0.05 → convert organically anyway | Recover 25% margin on 99,948 transactions | CRITICAL |
| Protect Organic Buyers | promo_dependency < 0.3 → never discount | Preserve ₹638/transaction premium | High |
| Target Persuadables Only | CATE > 0.15 → genuine incremental value | 62% CLV uplift on converted segment | High |
| Redefine KPIs | Replace gross revenue targets with net revenue + CLV | Align incentives with profitability | Medium |
Live Simulator
Reallocate ₹148M — Watch the P&L Recover
Based on your actual transaction data. Every number is traceable to a real calculation.
Sure Things + Hunters
60%
Budget % — low CATE segments
Budget wasted
₹89.2M
Persuadable Loyals
30%
Budget % — medium CATE
Productive spend
₹44.6M
Organic Buyers (Protect)
10%
Budget % — loyalty programs only
Premium retained
₹14.9M
Projected Net Rev/Txn
₹1,873
avg per transaction
Annual P&L Change
₹0
vs current strategy
Margin Recovered
₹0M
from CATE reallocation
Strategy Rating
😐 Current
Baseline
🎯 Current state: 60% of budget allocated to low-CATE customers. Net revenue locked at ₹1,873/transaction. Reduce "Sure Things + Hunters" allocation and shift to Persuadables to see P&L recovery.