Titulo Estágio
Predicting safe stock levels and optimal order sizes to avoid costly stock outs and overloaded warehouses
Local do Estágio
DEI / Incubadora do Instituto Pedro Nunes
Enquadramento
Since the opening of the first supermarket in Memphis in 1916, customers have come to expect that their desired products will be available on the shelves. However, this expectation is not always met. While consumers are disadvantaged by stock outs, these occurrences are even more critical for brands and retailers, resulting in significant financial losses, diminished credibility, and erosion of consumer trust. The high prevalence of stock outs—ranging from 18% to 24%—has transformed retailer data from mere sales and inventory records into valuable sources of information for forecasting. This shift has sparked a journey towards optimization based on historical records, marked by predicting future demand or the appropriate stock levels through machine learning (ML) techniques.
New ML models have added complexity to traditional time series models, which have proven to be fallible and inflexible. Indeed, ML approaches such as reinforcement learning, support vector machines, neural networks, and classifiers have been widely used for demand forecasting or predicting stock outs in complex models. However, these models often underperform due to errors in the recorded data by retailers and the inherent difficulty in capturing demand fluctuations. Recent evidence suggests that demand forecasting algorithms require manual adjustments to account for price changes, promotions, holidays, regulatory changes, linear stock outs, government policies, and competitor activities.
Brands and Ninjas has developed an algorithm already used by several clients, combining ML-modeled forecasts with contextual data provided by field auditors (our community of ninjas). Despite providing valuable insights for brands, the model is computationally complex, involving nearly a hundred predictors. Therefore, we have decided to revisit the conceptual approach of safe stock limits and delivery volumes, extensively used in the past, now leveraging new and more powerful ML models. This approach, while simpler, is easy to implement in-store as it allows for quick parameterization by the operator, indicating that when stock exceeds a certain limit, an order should be placed with a specified number of units.
Objetivo
With a three-year database from Delta Cafés, encompassing daily data for over 300 products across more than 400 stores, the objective is to determine safe stock limits that prevent stockouts and define order volumes to avoid both stockouts and overstocks. These limits will undergo blind cross-validation in a real retail environment or a simulated setting to verify their validity and reliability with future data.
Each record (row) in the database pertains to the performance of a product on a specific day in a particular store. The dataset includes daily sellout value, stock, shelf space, stock in transit, expected stock, and logistic route, among other variables.
Plano de Trabalhos - Semestre 1
1 - Literature Review of Machine Learning Models for Predicting Safe Stock Levels and Order Sizes
2 - Data Cleansing and Pre-processing
3 - Initial Exploration of ML Models and Preliminary Tests
Plano de Trabalhos - Semestre 2
1 - Production of ML Models and Performance Comparison
2 - Temporal Stability Testing and Ecological Cross-Validation
3 - Prepare Documentation for Communication with Retailers
4 - Writing the Final Dissertation
Condições
The work will be carried out at the Brands and Ninjas offices in IPN, located in Coimbra. Brands and Ninjas favor a hybrid work model, with the balance between in-person and remote work to be discussed.
Orientador
Pedro Bem-Haja
pedro.bemhaja@brandsandninjas.com 📩