Lenta Tech, the logistics domain of the Lenta retail group, utilizes demand forecasting to manage inventory levels across thousands of locations and product categories. The process involves balancing the risk of stockouts against the costs of overstocking, which is complicated by factors like promotions, seasonality, and supply chain disruptions.
The company employs two primary forecasting approaches. Simple models rely on historical sales data to identify trends and seasonality for stable products. More complex machine learning models incorporate external variables, such as pricing, promotional activity, inventory levels, and local traffic, to predict demand for thousands of items simultaneously.
A key challenge in forecasting is the 'censored demand' phenomenon, where zero sales do not necessarily indicate zero demand, but rather that the product was unavailable on the shelf. Because inventory decisions are based on previous forecasts, this creates a cycle where the model's own output limits the data available for future predictions.
Lenta Tech now focuses on generating a range of probable demand values rather than a single figure. This allows the business to align inventory levels with the specific economics of different product categories, such as prioritizing availability for high-margin items while minimizing waste for perishables.
The effectiveness of these forecasts is measured through business metrics, including lost sales due to stockouts, inventory write-offs, and turnover rates. The company emphasizes that forecasting is a continuous cycle requiring constant model updates and alignment between procurement, logistics, and marketing departments.
