13 / E-COMMERCE / AMAZON ADVERTISING & INVENTORY

A platform that watches every dollar of Amazon ad spend and every unit of inventory, and tells you exactly what to do about both.

Client: Amazon-based exporter

Tags: E-Commerce · Amazon PPC · Demand Forecasting · Optimization · MLOps

// Problem

• $1M/month GMV, scaling toward $2M+ — inefficiencies were set to scale right along with it
• 52.5% of SKU-days sat at zero available inventory
• 29% of PPC spend wasted — zero-sale campaigns plus runaway ACOS, up to 267% on some
• 6% repeat customer rate against a 20–25% category benchmark
• Launches and budget calls made on gut feel — no demand signal behind either

// The Contribution

A predictive analytics platform replacing five gut-feel decisions with five data-driven systems.

• PPC engine auto-pauses zero-sale campaigns and flags runaway ACOS before it burns budget
• Demand forecasting predicts stockouts 90 days out and triggers reorders against real lead times
• Inventory depletion is modeled with survival analysis, not a flat reorder point
• Return-risk scoring flags high-risk orders at the moment of purchase
• Every model is watched for drift and retrained automatically once its inputs move

// Results

• PPC waste cut 50–60% within 90 days
• SKU availability up from 47% to 70%+
• Repeat purchase rate roughly doubled
• $1.4M–$2.1M/year in recoverable value identified and sequenced against an ICE-scored roadmap
• Net contribution margin improved from ~26% to ~33.5%

// Technical Details

Anomaly Detection & Explainability

Isolation Forest over search-term-level spend, clicks, and impressions — every flag paired with a SHAP explanation, not just a score.

Demand Forecasting & Inventory Survival

Holt-Winters per SKU, weekly seasonality, confidence-bound forecasts with backtested accuracy. Kaplan-Meier and Weibull hazard models estimate time-to-stockout per SKU.

Optimization

LP solvers reallocate PPC budget across campaigns by expected ROAS. Markowitz-style portfolio optimization across SKUs. Inventory carrying-cost minimization under lead-time and capacity constraints.

Competitive Bidding Simulation

GSP auction simulator, Nash-equilibrium bidding solver, multi-round bid-war simulator — tests a bidding strategy against modeled competitors before any real budget is committed.

Drift Monitoring & Retraining

PSI and Kolmogorov-Smirnov tests catch feature drift against each model's training baseline. A drifted model is flagged and routed through a dedicated retraining pipeline automatically.