TL;DR
This paper proposes a data-driven framework to help supermarkets solve the "early morning replenishment" dilemma for fresh vegetables. By combining ARIMA time-series modeling, BP neural networks, and multi-criteria optimization, the authors provide a system to predict daily demand and set pricing strategies that maximize profit while accounting for the high perishability of produce.
What Problem Does This Paper Address?
Managing the "fresh food" section is a logistical nightmare for supermarkets. Vegetables have a razor-short shelf life—often less than 24 hours—and their quality deteriorates rapidly. Crucially, replenishment decisions must often be made between 3:00 AM and 4:00 AM, long before the day's actual demand or purchase prices are known.
The authors tackle the challenge of balancing inventory: overstocking leads to waste and "deteriorated appearance" discounts, while understocking results in lost revenue. The goal is to move away from gut-feeling replenishment toward an automated system that handles pricing (typically "cost-plus") and stock levels based on historical trends and seasonal fluctuations.
Methods
The researchers employed a multi-stage analytical pipeline:
- Data Preprocessing: They used the $3\sigma$ principle (three-standard-deviation rule) to identify and scrub outliers from a dataset of 1,095 entries, ensuring that extreme price spikes didn't skew the models.
- Time-Series Analysis: They established an ARIMA (AutoRegressive Integrated Moving Average) model to identify long-term distribution laws. This helped distinguish between seasonal trends (common in leafy greens) and long-term fluctuations (common in individual items).
- Correlation Modeling: To understand how different vegetables affect each other, they used Pearson Correlation Coefficients for broad categories and Spearman Correlation Coefficients for individual items.
- Demand Forecasting: A Backpropagation (BP) Neural Network was implemented in MATLAB to predict daily sales volumes for the upcoming week.
- Optimization & Decision Support: Finally, they built an optimization model to maximize the profit function: $\sum (\text{predicted sales} \times (\text{price} - \text{wholesale cost}) \times (1 - \text{loss rate}))$. They also used a Judgment Matrix (Analytic Hierarchy Process style) to weight the importance of external factors like supply chain reliability and customer feedback.
Datasets
The study utilized a dataset consisting of 1,095 days of sales data for six major vegetable categories (eggplant, cauliflower, pepper, aquatic rhizome, leafy vegetables, and floral vegetables).
- After preprocessing, 158 outliers (including zero-value entries) were removed, leaving 937 days of "normal" daily sales data for training and evaluation.
- The data includes unit sales prices, sales volumes (kg), and product codes.
Key Insights and Ideas
- Category vs. Item Behavior: The authors found that while broad categories like leafy vegetables and cauliflower show clear, predictable seasonal trends, individual items within those categories fluctuate much more erratically over time.
- The "Collocation" Effect: Using Spearman correlation, the study identified specific pairs of items that are frequently purchased together (e.g., specific codes for peppers). Strengthening these "collocation relationships" can help stabilize sales volumes.
- Information Hierarchy: Through their judgment matrix, the authors determined that supply chain data and customer purchase intention are the most critical factors for decision-making, outweighing even competitor pricing data.
- The 3:00 AM Constraint: The model specifically addresses the "blind" window of replenishment, providing a mathematical basis for decisions made before the market opens.
Key Results
- Model Fit: The ARIMA model achieved stationarity with a differencing order of 2, effectively capturing the "peaks and troughs" of vegetable sales.
- Correlation Findings: The heat maps revealed strong positive correlations between certain categories (e.g., solanacea and capsicum), suggesting that demand for these groups moves in tandem.
- Profit Optimization: The paper provides a sample "next week" strategy (Table 3), recommending specific expected prices and availability for items like Tremella and Chinese cabbage to maximize the objective profit function.
- Weight Ranking: The final analysis ranked the importance of data types as: Supply Chain Data = Purchase Intention > Purchasing Power > Customer Feedback > Inventory Data > Competitor Pricing.
Limitations and Future Work
- External Variables: While the judgment matrix acknowledges the importance of "special demands during holidays," the current quantitative model relies heavily on historical sales. It does not yet dynamically integrate real-time external shocks like extreme weather events or sudden local supply chain disruptions.
- Black-Box Forecasting: While the BP neural network is effective for short-term prediction, the authors note it functions as a complex mapping that may lack the interpretability of simpler statistical models.
- Future Directions: The authors suggest extending the framework to incorporate more granular external variables, such as weather patterns and specific promotional campaign data, to further refine the accuracy of the daily replenishment quantities.
Summary updated