PROBLEM (WHY? ): From our data exploration of the UPSERVE dataset, we noticed that there were quite a few anomalies. Anomalies could cost a restaurant a significant amount of money if the restaurant is unable to meet demand. How could we help mitigate the damage from these anomalies??
SOLUTION: Using time series anomaly detection, which mixes studentized residuals with Grubbs’ test, we will detect anomaly occurrences. Many anomalies do not just occur over one day. The idea is that if we can detect the occurrence of an anomaly the day after it happens, we can offer recommendations to the restaurant to help them capitalize on potential opportunities. (e.g. Nearby marathons)
USER INTERFACE: We implemented a local website UI that takes in a store id and then outputs a plot of sales revenue over a 30-day period, as well as whether the current revenue is an anomaly or not. If you are currently at an anomaly, the UI also outputs a list of events. (e.g. Nearby marathons)
WHERE UPSERVE COULD TAKE THIS: UPSERVE’s product currently gives restaurants percent changes in revenue. During our Datathon we did not have access to exact locations of restaurants. UPSERVE DOES have access to this and can combine that data with our anomaly detection model to give accuracy event listings and anomaly recommendations.