[ FORECASTING & DEMAND PLANNING ]

Demand Forecasting and Inventory Optimisation

Forecasting at SKU-store granularity is the hard version of forecasting. The long tail eats single-model approaches alive. Hierarchical forecasting with automatic model selection per series is the only approach that scales to thousands of SKUs without becoming impossible to maintain.

[ THE LEVENT POINT OF VIEW ]

Hierarchical per-SKU forecasting beats single-model approaches every time.

A single global model is comfortable to operate and wrong on the long tail. Per-store, per-SKU models are accurate and impossible to govern. Hierarchical forecasting is the middle path: a model framework per series with automatic selection from a small library, governed by a single pipeline. We have shipped it across multiple FMCG and retail engagements.

[ WHAT THIS MEANS IN PRACTICE ]

Our demand forecasting accelerator pairs LightGBM and Facebook Prophet, picks the better-performing model on rolling backtests per SKU-store combination, and orchestrates 12-month forward forecasts via Apache Airflow into Snowflake. A Tableau dashboard surfaces exception alerts only when actual demand diverges enough from the procurement plan to warrant attention. Buyers stop forecasting; they react to exceptions.

New product launches and promotional spikes are the classic forecasting failure modes. We handle launches with hierarchical priors borrowed from sibling SKUs. We handle promotional spikes with explicit promo features and a separate seasonality treatment. Neither is exotic, but doing both at scale is what the discipline is for.

Inventory optimisation is the natural next step. Forecasts become useful when they drive replenishment decisions: safety stock per SKU, reorder points per store, transfer recommendations across DCs. We layer a Mixed-Integer-Programming optimisation step on top of the forecast layer for customers who are ready to act on the outputs end-to-end. The forecast on its own creates a report; the forecast plus the optimiser creates a decision.

Data preconditions matter more in forecasting than in any other predictive workload. Two years of clean POS history at the store-SKU level is the floor; three years is comfortable. If the data is patchy or only available at the chain level, we say so before the engagement starts, and we either narrow scope (chain-level forecasts) or pause until the data foundation is right. The model amplifies whatever quality is upstream.

The handover to procurement matters as much as the model. We design the procurement workflow around the forecast outputs (monthly cadence, exception-driven attention) so the buyers actually use what the model produces.

[ HOW WE DELIVER THIS ]

How we deliver this

Build ships the forecasting framework and the orchestration. Operate runs the monthly retraining cadence and the model registry. Enable trains the buyer team on the new procurement workflow.

[ PROOF, NOT PROMISES ]

Accelerators that ship this in production today.

[ RETAIL & FMCG ]