Discount Spend Optimization

At a glance

In short	A modelling pipeline for an FMCG company that shows where trade discounts actually lift sales, and what happens to sales and profit if they change.
Problem	Discount budgets are spread across states, packs and channels, and it is hard to see which ones pay back.
What I built	One config-driven pipeline on Databricks for 18 product lines: data pull, merge, models, simulation and two dashboards.
How it works	Group segments with similar sales patterns, train one model per group, then change only the discount and re-predict to measure the effect.
Result	Separate hand-edited notebooks per product became one pipeline that runs for any product from an Excel config. Headline models reach R² 0.81-0.97; mature products 11-22% wMAPE.
Tech	Python, pandas, DTW clustering, XGBoost, SHAP, MLflow, Databricks notebooks and Apps, Azure Blob, Streamlit

One Excel workbook drives every stage: change a cell, not the code
Elasticity: predict sales at today's discount and ±0.5 points, then rate each segment Low, Medium or High
Saturation curves show where extra discount stops paying
Packs that take sales from each other get two models: one for the split, one for total sales
The P&L simulator loads the real models, so its numbers match the notebooks exactly
Every run is tracked in MLflow, with a model registered per cluster

Elasticity buckets

Bucket	Elasticity	What it means
Low	< 0.05	discount barely moves sales
Medium	0.05 - 0.5	some response, watch it
High	≥ 0.5	strong response, worth funding

Pipeline stages

Stage	Output
1. Data pull	one file per product per source
2. Merge	one model-ready table per product
3. Model	clusters, models, SHAP, saturation curves
4. Simulate	elasticity report and full what-if sweep
5. Dashboards	model results and P&L simulator

Technical details

Inputs: sales, discounts and price from the warehouse, retail audit, search share, household penetration, weather and a festival calendar
Segments are State × Pack × Channel by month, filled to continuous months so gaps do not distort the models
Features are picked per cluster by sweeping combinations; discount is always kept
The what-if sweep covers the 5th to 95th percentile of past discount in 0.1-point steps
Outputs are stored in Azure Blob by quarterly release, so only the stage that changed needs a re-run
Model quality is tracked per cluster with r², MAPE and wMAPE
The first version ran locally in pure pandas with Streamlit dashboards on Hugging Face Spaces

The code for this project is private, so there is no public link.