Retailers traditionally rely on sales trends, intuition, and various data streams to drive their product ordering. With today’s competitive environment, now more than ever supermarkets are investigating the applicability of superior technology like predictive ordering to provide a market advantage.
Predictive technologies are pervasive across many industries including, restaurants, insurance, telecommunications, banking, and health care. Retailers, in particular, stand to gain significant advantage given the potential to reduce waste and profit loss, along with generating higher levels of customer loyalty.
While AI/Machine Learning provides a better technology solution, we still need to start with the data. The multiple data streams include POS, pricing, loyalty, store door, assortment, placement, logistics, sell in, sell out, sell through, ship to, social media, competitive intelligence, and seasonality along with other external variables around events and weather.
Data can be combined in various ways to provide insight into future demand, for example, the success of a product launch. By analyzing planogram and assortment data and layering in point-of-sale and customer loyalty, predictions can be generated to better understand sales opportunities and thus increase order volumes. Understanding the balance between reliable SKU’s and new entrants or variants is key. AI-based models that better predict performance improve buying decisions and increase overall category volume, and remove bad actors before the numbers are irreversibly impacted for the financial period.
Rear-view mirror decision making can now finally be put behind us. With Machine Learning, (which is a subset of AI where we automate by using data and pattern recognition to enable applications to generate outputs without being explicitly programmed to do so), algorithms are used to build models that make useful predictions. In fact, most people interact with ML models on a daily basis through capabilities like shopping suggestions and targeted advertising. Pattern detection can help drive predictions regarding consumer behaviors, market fluctuations and much more. Our more historical predictive analytics, which have been provided by the various BI software vendors including our very own AFS G2, predict outcomes based primarily on historical statistical data applied through explicit programming, while Machine Learning is more of an evolutionary process making more accurate predictions over time as the algorithm learns by being fed more data.
Data utilization starts traditionally through descriptive analytics where we analyze the data to find out what happened. This is the traditional predictive analytics described above. Some vendors are putting a fancy voice-enabled query engine on the front end of this capability and calling it AI/ML, but what you really have is a voice-enabled dashboard running descriptive data queries or some type of search. With ML, once the data has been structured in a descriptive format, it can then be further evaluated with trends being identified and identifying why certain outcomes occurred. Machine learning as an extension of our more traditional predictive analytics looks at those trends and relationships, eventually predicting what will happen with higher accuracy over time.
Retail advantage can be gained by combining our traditional predictive analytics based on historical data with ML which is based on pattern recognition.
Filtering for the preferred minimum or maximum inventory level will provide insight into the average units sold during a specified time frame. We then calculate inventory days of supply that we have on hand which in turn provides a predictive indicator that will identify trends and through management by exception (MBE) we can then alert the appropriate individuals when action is needed.
Items generating higher volume sales analyzed at a store-by-store level over time can uncover trends in consumer behavior and preferences that will provide insights on what items to carry and at what price points. These types of predictions lead to better decision making in ordering, stocking and other factors affecting the company’s bottom line.
Comparing sales and inventory for similar periods of time by geographic region including influencing factors such as competitor brand actions can help to determine adjustments to orders and inventory levels.
Further enhance order upside
SALY, or same as last year, seems to be the dominant order pattern prediction and inventory level planner used for major brands in the market. However, the demand environment can shift for multiple reasons such as something like an uptick in residential growth profiles. This method also does a very poor job accounting for unmet consumer demand.
Machine Learning techniques will move beyond SALY and look, for example, at a 60-day sales velocity run rate leading up to the time period being analyzed for both last year and this year. If this year’s sales velocity is stronger, that pattern leads to a prediction of higher sales and that a higher on-hand inventory level is needed to capitalize on that sales opportunity. Highly ranked attributes affecting demand will then be added to this baseline pattern which is what drives higher levels of accuracy. Tracking sales and inventory trends over time and analyzing the period-over-period results is core to predicting consumer needs and to order/stock accordingly.
Machine learning pulls from multiple factors
Consumer demographics, various store attributes and weather can provide additional inputs to ML algorithms that then drive deeper trend analysis and predictions. Data elements such as these empower the machine learning engine to predict pattern changes in sales, providing reliable evidence to suggest and predict the required order volume to capitalize on this demand. Communicating the new inventory requirements upstream in a timely fashion will avoid exposure to excess or insufficient stock levels, thus improving efficiency and generating increases in gross margins and ROI.
Of course, as we all know, machines make mistakes. AFS has structured a set of dashboards and workbench’s as part of our solution framework to enable human interaction as part of the decision making process. Everything needs to be kept transparent.
The predictive order solution framework
Given the data and variables involved in predicting potential order upside with high accuracy and providing the solid suggested order profile optimizing targeted KPI’s, while at the same time limiting any downside related to inventory exposure, the AFS solution framework is broken down as follows;
Suggested and predictive orders
Saving time, our most precious commodity is critical when it comes to in-store activities. With ML-based predictive orders, we have seen improvements of 10 minutes per order with associated benefits in improved order accuracies. Taking the guesswork out of shelf management, procurement, and customer engagement and ultimately knowing what to do and when to do it is a significant improvement as compared to today’ practices.
Sales opportunity detection and creation
Knowing the buying behavior and assortments of similar store doors in comparable geo demographic regions empowers field personnel to identify upsell opportunities in their markets. This capability and much more is created through our Machine Learning algorithm which models buying behavior trends, analyzes transaction activity, and generates new sales opportunities by providing suggested orders to the field.
Shelf trend prediction
Understanding trends that impact shelf positions requires a varied set of data inputs including the usual transactions, inventory, and promotions. In addition, variables captured through social listening, sentiment, and preference now play an increasingly important role as part of the data set. The purpose of applying Machine Learning in this area is to update the forecast with more reliable predictions on a near real-time basis. Previously undetected patterns would be brought to light and appropriate actions are prescribed through the ML algorithm. Strong results have been recorded thus far across sales, out of stocks, overstocks and overall assortment efficiency.
We can target pattern shifts that will indicate potential churn where a customer is shifting demand to a competitor. These declining buying patterns can be picked up by removing dips based on seasonal patterns and then modeled so that we eliminate false recommendations that are common due to the nature of today’s forecasting algorithms.
Promotion lift prediction
Both short and long term planning requires accurate forecasting. Efficient resource allocation is dependent on the accuracy of this forecast. The difficulty lies in the sales data itself where the combination of high variability along with the complexity of the predictive relationship between variables make it too complex to solve with high accuracy using traditional methods. Given that interaction patterns between variables are not static, traditional linear modeling techniques will prove to be inefficient in some cases and thus machine learning is a better approach to drive increased accuracy. Our current Machine Learning model considers a number of variables including key influencers such as length, tactics, price, discounts, pack size, competitor actions and shelf position. Based on the variables the model will calculate uplifts in terms of accurately predicting actual overall sales volume as compared to baseline.
Performance gaps in the execution of brand and category strategies require immediate attention given the implications related to dependent processes at the S&OP level. Areas like spend effectivity and potential impacts to spend including competitor actions are key variables considered in the machine learning algorithm.
360 revenue insights
Driving revenue insights requires a single version of the truth across multiple sources of data. This model would provide transparency into the complex and inconsistent hierarchies across your manufacturing, distribution, and retail channels. Machine Learning is evolving rapidly in this area as data is made available.
Traditional methods in this area involve various forms of explicit programming which includes various types of LP-based optimizations. While this approach has been useful for many years, there is a significant amount of bias which greatly affects the accuracy of the predicted results. With Machine Learning the data can now do the talking as we translate corporate strategies into optimized plans for all customers and channels in order to grow both brand and category.
An important aspect of the move toward AI/ML is that the technology isn’t replacing what we do as humans. In fact, AI/ML is doing what we are not capable of doing such as analyzing enormous amounts of data thus freeing up our time to do what we do best such as negotiating deals.
In its Technology Vision for Consumer Goods Report, Accenture reports, “Seventy-eight percent of industry executives agree that AI will revolutionize the way we gain information from and interact with customers.”
And it’s going to be everywhere. “It’s weaving its way into nearly everything,” says Gary Hawkins, CEO of Los Angeles-based Center for Advancing Retail & Technology (CART). “Retailers have realized the world is changing really fast and they’re not competing just with the store down the street but with Amazon, Walmart, Kroger, who are investing tens of millions of dollars in technology every year.”
Essentially the AFS Predictive and Suggested Order solution platform delivers on answering key questions such as which products should I promote, what order mix will drive the highest revenue/margin per square foot, what prices should I charge and what inventory should I maintain in-store? AI/ML technology continues to learn over time, so it gets better and better at knowing the answers to all these important questions.
Personalization is very near term as stores could soon know who’s walking down their aisles enabling us to recognize a person’s dietary preferences, the recipes he or she likes and any food allergies, then use that information via mobile technology to recommend various products or to show what’s on sale. Smart-shelf technology can also play a big role in this area. Sensors can detect individuals who opt into the in-store application and send personal pricing and product selections to them, via their mobile device and shelf displays. Consumer order placement could become somewhat automated where AI would predict a shopper’s grocery order and suggest what they need.