Notion Demand: Building a Retail Inventory Forecasting Model from Scratch

Utsav joined Notion Demand in June 2024 as one of the first researchers on the team. The mandate was broad: explore how AI could optimize inventory management. His job was to turn that into something buildable.

His first step was to define the scope. He evaluated which industry would yield the richest data and the clearest problem, landing on retail because of the volume of SKU-level data available and the well-documented demand variability. He selected the M5 Walmart dataset — a publicly available, large-scale retail sales dataset — as the foundation for the analysis.

Exploratory Analysis and Hypothesis Testing

Utsav ran a full exploratory data analysis on the dataset, identifying demand patterns tied to time-of-month cycles, holidays, government subsidy disbursement dates, and product category differences across food, health, and frozen goods. He segmented the data by city and category, then ran hypothesis testing to validate which patterns were statistically meaningful.

The CEO reviewed progress twice a week, and at each gate Utsav presented the reasoning behind each narrowing decision: why retail, why this dataset, which KPIs were holding up under testing, and what the model would need to do.

Building the Model

For the forecasting model, Utsav designed a hybrid parallel architecture that stacked multiple model types — LSTM, GRU, linear and non-linear regression, CNN — in parallel, feeding their outputs through an attention layer and a neural network to produce the final forecast. The architecture was designed to allow complexity layers to be switched off, enabling latency optimization across different time horizons and use cases.

The full sprint ran two months: roughly two weeks on literature review, one to two weeks on data selection and EDA, one week on hypothesis testing, and two weeks on model building. He delivered a working MVP and full documentation.