Lucas Reymond

While working at Panipro, I developed a forecasting pipeline for predicting future product demand across bakery stores. Historical sales provided a natural training target, but they did not always represent actual customer demand: when a product sold out before closing, subsequent demand was unobserved. Inventory discrepancies and cashier-entry errors introduced additional noise.

I developed a statistical method to estimate this latent, unconstrained demand from timestamped sales and inventory events. For each product and day of the week, the method learns a typical cumulative intraday sales profile using days without detected stockouts. When a stockout occurs, it uses the time of the last sale to estimate the missing portion of the day’s demand. A regularized estimator combines the observed sales with the product’s historical baseline, limiting the amplification of noisy transactions.

The pipeline includes conservative safeguards for insufficient history, unusually early stockouts and extreme estimates. It produces corrected daily demand values together with stockout diagnostics, estimation methods and confidence indicators. I also built interactive visualizations to compare observed sales with estimated demand for each product. The resulting data provides a more representative target for demand-forecasting models than raw historical sales alone.