Discounted Sales of Expiring Perishables: Challenges for Demand Forecasting in Grocery Retail Practice

Grocery retailers frequently apply price discounts to stimulate demand for expiring perishables. However, integrating these discounted sales into future demand forecasts presents a significant challenge. This study investigates the effectiveness of incorporating a fixed share of these sales as \textit{regular} demand into the forecast, as commonly applied in practice. We employ a two-step regression approach on data from a major European grocery retailer, covering over 1,700 products across 676 stores. We reveal that forecasts underestimate actual demand for most SKUs when discounted sales occur. This residual uplift effect is significantly influenced by the number of sales at reduced prices. Our findings underscore the necessity for more precise approaches to integrate discounted sales into demand forecasts, thereby preventing excess inventory and the associated economic and environmental impacts of spoilage in the grocery sector.

💡 Research Summary

The paper investigates a practical problem faced by grocery retailers: how to incorporate sales generated by price discounts on expiring perishable items into future demand forecasts. Using a rich dataset from a major European grocery chain, the authors analyze 1,705 SKUs across 676 stores over 92 business days (August‑October 2024). Each daily record contains beginning‑of‑day stock, the retailer’s internal forecast, actual sales, and the number of units sold at a reduced price. After filtering for SKUs with sufficient activity (≥100 days of data and ≥50 days with discounted sales), the final sample comprises 1,132 SKUs and roughly 20 million observations, of which only about 3 % involve discounted sales.

The authors adopt a two‑step linear regression framework. In the first step they estimate a baseline demand model using only days without any discounted sales. The model regresses daily sales on the retailer’s forecast, the available stock, and weekday dummies. Ordinary Least Squares (OLS) yields fitted values for “non‑discount” days; the residuals (actual minus fitted sales) represent the forecast error under normal conditions.

In the second step they focus on days with at least one discounted sale. Using the same covariates plus the count of discounted sales (DS), they regress the previously computed residuals on these variables. The coefficient γ̂₁₀ on DS captures the marginal “uplift” effect: how much each additional discounted unit increases the residual (i.e., the amount by which the forecast under‑estimates actual demand).

Key findings are striking. For 92.2 % of the SKUs the average residual during discount periods is positive (mean ≈ 0.41 units), indicating systematic under‑forecasting when discounts are applied. The uplift coefficient γ̂₁₀ is positive for 1,067 SKUs, and statistically significant at the 5 % level for 1,044 of them. No SKU shows a significant negative uplift, suggesting that discounts never depress demand in this context. The histogram of γ̂₁₀ values shows a concentration around a similar magnitude, yet with notable variation across products, implying that a one‑size‑fits‑all rule (e.g., treating a fixed share of discounted sales as “regular” demand) is inappropriate.

The paper discusses practical implications. Retailers often assume a fixed proportion of discounted sales is independent of the promotion and feed that share into future forecasts. The empirical evidence shows this practice leads to systematic under‑prediction, which can trigger unnecessary additional discounts, higher inventory levels, and ultimately more food waste. Because the uplift effect varies by SKU, the authors argue for tailored forecasting approaches that explicitly model the promotional impact.

Limitations are acknowledged. The dataset lacks information on remaining shelf‑life, the number of items bearing discount stickers (as opposed to actual discounted sales), and other contextual variables such as store location, holidays, or the presence of substitutes/complements. The linear regression framework cannot capture non‑linearities or interaction effects that may be present. Consequently, the authors propose future work employing more advanced machine‑learning techniques (random forests, neural networks) and richer feature sets to improve accuracy and to estimate heterogeneous treatment effects (e.g., using causal trees).

In conclusion, the study provides the first quantitative assessment of the “uplift” caused by discounting expiring perishables in a real‑world retail setting. It demonstrates that discount‑driven sales substantially increase actual demand beyond the retailer’s baseline forecast, and that the magnitude of this effect differs across products. These insights call for more sophisticated, SKU‑specific demand‑forecasting models to avoid over‑stocking, reduce the need for repeated price cuts, and mitigate the economic and environmental costs associated with food spoilage.