An Interpretable Decision-Support Architecture for Batch Production with POS-Only Operational Data: A Point-in-Time Evaluation

Small bakeries have to decide how much bread they will make in discrete batch units based on short and irregular sales histories. Also, point-of-sale (POS) logs have a clear limitation. After a bakery sells all the bread it made, we cannot track how many customers actually tried to buy later. Only paid sales are recorded. In this research, we evaluated RSB Forecaster, a decision-support system which helps them decide the amount of bread to make. This system uses Median-14, a model that takes the median of the last 14 observed open-day records and adjusts it by day of the week. To recommend final batch counts, this system combines the Median-14 prediction value with three factors: proxy batch sizes fixed from past POS records, uncertainty intervals calculated from prior-fold errors, and a rule-based review gate that picks items for a person to check. When we evaluated Store A, for 182 days of that data we predicted the next-day amount of bread sold by moving the forecast date forward one day at a time. Here, we only used two factors from outside: day of the week and holidays. To prevent data leakage, we excluded all information that we cannot know on the forecast day, such as target-day actual weather, closed days guessed from future POS sales, and full-period metadata, and we did not use prediction errors (residuals) from the same evaluation fold. Across all 4,810 date-SKU rows targeted for evaluation, the mean absolute error (MAE) of the prediction value was 1.87 units and the weighted mean absolute percentage error (wMAPE) was 60.09%. The recommended amount of batches matched the sales-based proxy batch count exactly on 57.90% of rows, and fell within 1 batch on 90.44% of rows. The 80% uncertainty intervals built from prior-fold residuals covered actual sales on 82.31% of the 4,782 eligible rows. The review gate classified 63.11% of eligible rows as 'human review'. Auto-passed rows had better within-1-batch accuracy than flagged rows (95.58% vs. 87.44%). The reason why auto-passed rows had better accuracy is that they had more rows with a zero rounded proxy-batch target, which is easy to predict. Also, we compared Median-14 one-to-one with 20 different machine learning regression models on the same rows. From that result, Median-14 had no statistically significant difference from Huber regression and LinearSVR, and performed better than the other 18 models. The same tendency appeared in Store B, which had a 1,422-row fixed-cutoff stress test. While no rolling-median window length won first place for every test metric, the attempt to adjust for stockouts using rules based only on POS data actually dropped the accuracy of the prediction. Because our dataset does not have records of actual production, inventory, sell-out times, waste, or manager actions, this evaluation is limited to observed sales and simulations on past data. The main contribution of this study is to provide a reproducible way to audit next-day batch production recommendations while controlling data leakage.