Home/Work/ Patanjali
CPG · India · AI & Forecasting

Forecasting that
survives volatile markets.

Ensemble ARIMA + gradient-boosted model with a model-monitoring layer. Built for Patanjali Ayurved to reduce commodity hedging cost and accelerate procurement planning cycles across 12 raw-material SKUs.

ClientPatanjali Ayurved
SectorCPG · India
Duration10 weeks (shipped 3 weeks early)
PracticeAI & Machine Learning
−18%
Hedging cost reduction
Faster planning cycle
4wk
Reliable MAPE forecast window
The challenge

Volatile markets. Manual planning. Expensive hedging.

Patanjali Ayurved sources 200+ raw materials — herbs, carrier oils, packaging materials — many of which are priced on commodities markets that can swing 20–30% seasonally. Their procurement team was hedging conservatively (expensively) because their planning horizon was too uncertain.

The existing process relied on spreadsheet-based 2-week rolling averages, with manual adjustments for festivals and monsoon cycles that required senior analysts to spend most of their week on forward-looking planning rather than strategic work.

S.T.R.In.G was engaged to build a forecasting system that could reliably predict demand and price movements 4 weeks out — enough lead time to hedge more precisely and negotiate better supplier contracts.

Our approach

Ensemble modelling with model health monitoring.

Phase 01 · Weeks 1–2

Data Audit & Feature Engineering

We audited 3 years of transaction data, identified data quality issues in ~12% of records, and engineered 40+ features including local festival calendars, monsoon indices and commodity futures as external regressors.

Phase 02 · Weeks 3–5

Model Development

Trained ARIMA, Prophet and LightGBM models independently. Evaluated against a held-out 8-week test window. Blended predictions using a meta-learner that weighted models by their recent accuracy on each SKU.

Phase 03 · Weeks 6–7

Model Monitoring Layer

Built drift detection on both input features and predictions. Automated weekly retraining pipeline. Alerting when a SKU's rolling MAPE exceeds threshold — surfaced to analysts before hedging decisions are made.

Phase 04 · Weeks 7–8 (3 weeks early)

Integration & Handoff

Forecasts pushed directly into Patanjali's procurement planning tool via API. Analyst dashboard showing forecasts, confidence bands and explainability drivers. Runbook and 2-day handoff training for internal team.

The outcome

Shipped early. Adopted immediately.

The ensemble model delivered 4-week forecasts with a MAPE consistently below 8% across all 12 production SKUs — significantly better than the 18–22% error of the previous spreadsheet approach. The procurement team began using the output in week 9, before the project officially closed.

Hedging cost fell by 18% in the first full quarter of operation as narrower confidence intervals allowed the team to take smaller, more precise hedging positions. Planning cycles that previously required a fortnight of analyst time were cut to 4 working days.

"We expected to spend six months validating this. We were using it in two." — Procurement Head, Patanjali Ayurved

The model-monitoring layer has now flagged three significant market-structure shifts, triggering targeted retraining before forecast accuracy degraded — exactly as designed.

Tech used
Python ARIMA / statsmodels Prophet LightGBM MLflow Apache Airflow FastAPI PostgreSQL AWS EC2 / S3
← All case studies Work with us on something similar