CASE STUDY
Predictive Demand Forecasting on Google Cloud AI Cuts Inventory Waste by 30% for a Regional Retail Chain
Key Results
| Inventory Waste | 30% reduction in inventory waste within the first two quarters post-launch |
| In-stock Rate | 18% improvement in in-stock rate across 50+ locations |
| Forecast Accuracy | Accuracy improved from 61% to 87% against actual sales |
| ROI Timeline | Full investment recouped within 9 months of deployment |
SERVICE AREA
INDUSTRY
CHALLENGE
Inventory decisions running on instinct at scale — chronic overstock, stockouts, and eroding margins
A large retail organization with many store locations (over 50) has an unresolved demand forecasting issue that it has been covering up through extra safety stock and buyer intuition. Purchase decisions are based on past year sales, buyer experience, and outdated spreadsheets. Consequently, the company is consistently experiencing chronic excess inventory for its slower-moving products, intermittent out-of-stock conditions on fast-selling SKUs during peak time frames, and, as a direct result of markdowns from inventory that should never have been purchased, reduced profit margins.
All necessary data exists: years of Point of Sale (POS) history, vendor lead times, promotional calendars, and seasonal trends – but the infrastructure to utilize this data does not exist. Data resides within disparate systems, transformations occur manually, and there is no methodical approach to provide historical context to purchasing decisions on a scale that will allow the retailer to grow. The growing gap between what can be determined by utilizing historical and real-time signals across all locations and SKUs simultaneously and what current buyers know will continue to expand unless some type of predictive analytics system can provide actionable recommendations before each buying cycle.
SOLUTION
Custom AI-powered demand forecasting system built on Google Cloud with SKU-level precision and full ERP integration
SCIGON developed a customized AI-based demand forecasting solution on Google Cloud that utilized client data (from POS, ERP, and Vendor Management Systems) within a single, unified data stream created using BigQuery and Cloud Dataflow. The development process began with the data engineering phase. SCIGON’s data engineers cleaned up years of poorly formatted point-of-sale (POS) data and standardized the data format, as well as created a historical baseline for all their SKUs. This historical baseline provided a foundation on which the machine-learning forecasting model(s) could be trained. In terms of the forecast, the SKUs were forecasted individually. However, each SKU had different variables to account for, such as seasonality, location-based factors, promotional lift, etc. These variables impacted how the demand for each SKU was modeled. Once these models were trained, they were validated by comparing the predicted values of a holdout period of historical data to the actual values during the same time. The results of these validations determined whether the models would be deployed. Additionally, once the models were deployed, their performance would continue to be evaluated against actual sales.
The results from the models were also made explainable; when a buyer received a recommendation from the system regarding what to purchase in terms of an amount, it also showed them why they should purchase those items — i.e., what variables influenced the decision. This increased the rate of adoption for the system by allowing users to understand how the system arrived at its conclusions. Ultimately, the system was incorporated into the current buying workflow of the client. Instead of having to utilize another interface to access recommendations, buyers would receive recommendations in the interfaces they currently use. As a result, the company now has a reliable way of determining exactly how much of any product it needs to purchase, when to do so, for which locations, and how many products are needed to fulfill customer demands.
