For security integrators who deal in the retail vertical, point-of-sale (PoS) systems generate rich data, which is a resource being used by such customers for exception reporting.
This same data can support many other advanced analytic applications, as well. For instance, research within Stanley Security is underway using PoS data to predict whether promotions will be successful and result in sales lift of products.
In a recent experiment developed by the Stanley Security Insights team, PoS data was analyzed to determine the effects of promotions on sales, and, more accurately, forecast the effects of future promotions on sales and inventory.
A predictive model was created to forecast sales during promotions and price changes. In tests of the model, 77% of variance of sales was explained for a product. Thanks to improvements outlined below, that number is expected to increase even further.
During the exploration of the data, the ability to model price elasticity was discovered: this model has the potential to optimize price toward maximum sales. Potential next steps are optimizing price across products, and optimizing price to maximize margin.
In addition, interactive dashboards were created using a visualization tool that enabled the end user to filter data by customer, store, item as well as promotion. The visualization tool also has the ability to send notifications to an assigned user, so that quick action can be taken, such as inventory falling below a threshold level.
More accurate sales predictions can lead to improved supply chain planning to ensure customers’ shelves stay appropriately stocked; and improved promotions planning is achieved through more accurate predictions.
Value Propositions Include Improved Promotions Planning
The potential business value propositions are numerous. For starters, they include improved promotions planning so that channel marketing and finance can more precisely predict ROI when writing and approving price change requests.
Understanding the combination of sales lift and margin can help identify the most profitable promotions. Visibility to performance can help optimize promotion duration, type, time period and price points.
Supply chain optimization is also a consideration whereby demand planners can more accurately anticipate the lift associated with similar past promotions, driving forecast accuracy and potentially reducing “out of stock” scenarios.
Sales can utilize data to sell additional product to customers based on history. Pricing optimization is another area to maximize sales. With margin data, price can be optimized to maximize profit.
Price can also be optimized across products. (The price of hand drills in relation to each other affects sales of each item, for example.)
Current Model Influenced by External Variables
The primary data sources are PoS data, promotions data and external data sets available for free in the public domain. A regression model was created to explain change in unit sales.
This particular regression takes the log of unit sales and the predictor variables, which consist of price of item, seasonality by month of the year, active promotion for item, macroeconomic data — S&P 500 price, price of product and substitutes.
Of note, the use of external variables such as LIBOR (London Interbank Offered Rate) and the S&P 500 were found to be powerful predictors in the model, presumably because they are a proxy for the economy.
Various machine learning models were used including random forest and regression. Both yielded results illustrating that the model explains more than 70% of the variance in sales.
Myriad Potential Model Improvements Available
The model outcomes can be improved by adding the following:
- More years of historical PoS data
- Promotions bible of past years
- More accurate MSRP unit price data
- Competitor MSRP prices
- Advertising and marketing efforts (by total dollars and by channel)
- More comprehensive macroeconomic data
- Data on the health of specific retailer
- Machine learning techniques (random forest, etc.)
- Margin data to optimize for profit
- Optimization models across products (pricing of substitutes)
User Interaction Can Be Customized
In addition to predictive use case, it was discovered that exploratory visualizations of the PoS data were useful to organizations. For example, one dashboard shows sales by customer, store and item. It also details the percentage of weekly change of sales.
A second dashboard shows inventory on hand for retail locations filtered by customer, store number and item. A red bar indicates whether balance on hand is below a preset threshold value.
PoS data offers a deep resource, particularly when combined with other data such as promotions numbers. Future research will explore market basket analysis and pricing elasticity, as value-add for retail market integrators continues to virtually sell itself.
Dave Bhattacharjee is Vice President, Data Analytics, for Stanley Security. Reach him at [email protected]