Information is being consumed and shared through social media channels at lightning speeds—effectively helping or harming brands on a daily basis. Social media platforms, such as Facebook, Twitter and Instagram, are forcing companies to redefine e-commerce, marketing, and operational strategies and look for new ways to maximize the impact of this new consumer technology on business. But how is social media impacting loss prevention?
The Retail Equation recently conducted a study to determine if it was possible to predict retail shrink for a given store from a collection of tweets originating from locations surrounding that store. By predicting shrink, we are able to create a geographic risk index in all areas (even beyond the areas with stores in the sample).
To conduct the study, we received a test feed from Twitter that included between 20,000 and 100,000 tweets per day. Over the course of the study, more than 10 million tweets were collected. We identified tweets that occurred near a retail location and converted these tweets into metrics that could be used to determine tweet sentiment, presence of key topics of interest, and then convert these words into meaning in a mathematical space.
Tweet Sentiment and Topic Mapping
Beyond content, every tweet has an overall “sentiment.” Sentiment scores are based on the words in a tweet and result from a relationship of those tweets to emoticons. The scores are assigned to each word on a scale from very positive to very negative.
Additionally, tweets are converted into mathematical vectors to allow us to categorize the tweet and measure the concentration of various topics around each store.
To process the 10 million tweets, we used the MapReduce tool on a Hadoop infrastructure to speed up the time needed to sift through all the data. Tweets were cleaned for spell check, stemming (reducing a word to its root form, like going from “sitting” to “sit” for easier language processing), and removal of stop words (common words like “and,” “the,” and “is,” which can cause problems in text mining). We then counted word frequency, distance between a tweet and each store, and topic frequency metrics. This allowed us to create an average sentiment for each store.
Tweets Do Predict Shrink
Once all tweets were analyzed, we were able to draw statistical conclusions that illustrated if overall Twitter sentiment surrounding a store is negative, the store exhibits poor shrink metrics. Like crime reports, it is a truly external tool, outside of the internally focused metrics that retailers typically collect in their charge to better understand shrink. But unlike crime reports, it is far more broadly based on the voice of the shopper and, since it is social-media based, it is much timelier.
There are many ways we expect this knowledge to be used. Twitter data can be combined with point-of-sale data and other risk indices to create a detailed shrink prediction for each store. It can also be implemented within existing solutions—return authorization, exception reporting, online CNP tools—as a new big data feed to improve risk scoring.
There are many more possibilities to explore, but the key takeaway from this research is this—social media can be a powerful tool that retail loss prevention can, and should, exploit.