An Application of Marketing Analytics to Estimate the Impact of Out-of-Stock Instances for Retailers
Sales figures often fail to accurately reflect the true demand observed by retailers. While sales data provides insights into consumer purchasing patterns, it does not capture the full picture of consumer behaviour. There are several reasons for this discrepancy. Firstly, sales figures only consider completed transactions and do not account for potential sales lost due to stockouts, pricing issues, or poor customer service. These missed opportunities significantly impact retailers' understanding of the actual demand for their products.
Furthermore, external factors unrelated to consumer demand can influence sales figures. Seasonal fluctuations, promotional activities, and sudden market trends can artificially inflate or deflate sales numbers, leading to an inaccurate representation of true demand. For example, a retailer may experience a sales surge during a specific holiday period, but it may not indicate a sustained increase in demand throughout the year. Relying solely on sales figures can result in misinterpretation and a misalignment with customers' genuine needs and preferences.
When retailers face uncertain demand, they can utilize observed sales to update their demand estimates. However, this learning process is constrained by the inventory they carry. When demand exceeds inventory, such as during an out-of-stock event, retailers generally cannot directly observe the actual demand. Accurate demand forecasting becomes crucial not only to reduce out-of-stocks (OOS) but also to minimize overstocking (OS). This use case focuses on a food retailer operating through multiple franchisees, aiming to provide customers with restaurant-quality foods, great value, and unique service. Their product lineup showcases top-quality, exclusive, flash-frozen items that are not available in other food retailers.
To estimate out-of-stocks (OOS), it is necessary to know the inventory levels at the store and the demand experienced by the store. In this case, the supplied data includes end-of-day inventory on hand, as reported by the stores to the head office. However, it is important to note that some records may be blank or contain negative values, which were eliminated from the analysis. Figure 1 provides an overview of the specific data fields used in the predictive models for reference. Eliminating these from the analysis cut approximately 1% of the records. In Figure 1, we present the data fields are used in the predictive models:
Figure 1. Data Fields used in the Predictive Model
The newly developed daily model has shown impressive accuracy in predicting sales, with an average percentage error of just 2% at the day and store level. This level of precision is a significant improvement compared to the forecasts generated by the retailer itself. To provide context and demonstrate the superiority of the new model, Figure 2 presents a comparative analysis of the accuracy between the two forecasting approaches. Both forecasts were generated six weeks prior to the commencement of the promotion period and flyer distribution.
By evaluating the performance of the new model against the retailer's forecasts, it becomes evident that the new model offers a more reliable and robust prediction of sales. The new model was more accurate in every Price Look Up (PLU), a KPI equivalent to SKU, in the test; on average with a WMAPE of 7.5% vs the retailer’s mean absolute value weighted by store sales of 32%. The low average percentage error signifies that the new model is consistently close to the actual sales figures, enabling the retailer to make more informed decisions regarding inventory management, resource allocation, and overall business strategy. This level of accuracy is particularly valuable when preparing for promotional activities and flyer drops, as it allows the retailer to anticipate customer demand and ensure optimal stock availability to meet the expected surge in sales.
The utilization of the new model, with its significantly improved accuracy, empowers the retailer to proactively respond to market dynamics and enhance operational efficiency. By having access to reliable sales forecasts well in advance, the retailer can make timely adjustments to their supply chain, optimize pricing strategies, and allocate resources effectively to maximize profitability during the promotion period. The superior accuracy of the new model instills confidence in the retailer's decision-making process, ultimately leading to better customer satisfaction, increased sales, and a stronger competitive position in the market.
Figure 2. Forecast Accuracy Comparison
The next step is to estimate the OOS losses. For our analysis, only the days in which the PLU is on promotion, with a discount greater than 0 are used. As a conservative approach, a predictive model was used to first group store-days into deciles based on their expected sales. As far as the model is concerned, sales for store-days within a decile should be virtually the same. Then, the stores are divided into two groups: a high inventory group, which has inventory on hand at much higher levels than the predicted sales, and a second group in which inventory on hand is much closer to, or even less than predicted sales. Then the difference is taken between the actual, observed sales for both groups as the effect of having higher inventory levels on hand. In other words, where stores order much more inventory than they are likely to need, we expect sales to be higher than if they have little excess inventory on hand or if they run out of stock.
In Table 1, we show a set of store-days for PLU 30, Italian Meatballs. The model predicts sales for these stores will be very similar 9.7 and 9.6 units per day. However, the Hi in Table 1, we show a set of store-days for PLU 30, Italian Meatballs. The model predicts sales for these stores will be very similar 9.7 and 9.6 units per day. However, the Hi Inventory group has, on average, 7.7 units on hand before the start of the current sales day (i.e. yesterday night) while the Low Inventory group has only 1.5 units.
Table 1
In theory, the only difference between the two store-day groups is inventory. In all other respects, they should generate the same sales, but they don’t. The Low Inventory group generates 2.3 units per day less. The Hi Inventory group generates 10.1 units, whereas the Low Inventory group generates only 7.8 units. This difference is aggregated across all store-days and multiplied by the sales price of the day to arrive at an estimate of dollar loss. When we divide this dollar loss by the annual actual sales, we arrive at the percentage figures for 5 specific PLUs presented in Table 2 below:
Table 2