Cornershop: how machine learning can improve customer satisfaction and increase accuracy in operations
Through neural networks Cornershop has improved the accuracy of the estimations of delivery time of products.
Cornershop: delivery of groceries to your front door
Cornershop is a Chilean startup that provides on-demand home delivery of groceries. Customers buy products from the Cornershop mobile app and receive the products the same day [1]. The app was launched in May 2015 in Chile and in July of 2015 in Mexico with an initial investment of $1 million.
Machine learning at Cornershop
In the online retail business, delivery times is a key indicator that is reviewed constantly. In the case of Cornershop, a better estimation of delivery time estimates improves customer satisfaction and enable a better assignation of the delivery staff to the different purchase orders. The purchase process at Cornershop (Figure 1) relies on many factors that increase the variability of delivery times. The incidence of traffic and the distance between the location of the customer and the supermarket affect the transportation times (stages 3 and 5). Correspondingly, the mix of products ordered and their location in the store, the skill level of the delivery person and the congestion in the stores affect the speed of purchase (stage 4).
Figure 1. Purchase process at Cornershop.
Machine learning has emerged as a possible solution to improve the delivery time estimate. When Cornershop began its operations, the logistics team predicted the delivery time based on historical purchases with similar characteristics [2]. However, due to the large number of factors that affect the efficiency of the delivery process, the team decided to apply more sophisticated methods and train a model for learning from previous experiences.
What the management team is doing and what it will do to improve the prediction of delivery times
In the short term, to improve the prediction of the delivery times and the assignation of delivery staff [3], Cornershop has strengthened the data science team and has started using machine learning methods. To illustrate, the company applied four machine learning methods. For the predictions of each method, the team calculated the standard error of the mean (SEM) and the mean absolute error (MAE). SEM measures the deviation of the predictions considering their directions and therefore positive deviations can offset negative deviations. In contrast, MAE measures the deviations of the predictions without considering their directions.
The method that obtained the best results was the neural network, which is a technique that simulates the neurons of the human brain through the creation of multiple layers that process information with the objective of transforming the input into something the output unit can use [4]. Through this method, the MAE was reduced by approximately 40% [2] versus the non-machine learning method. Since positive deviation canceled negative deviation, the SEM results were similar in both cases.
While Cornershop´s operations were improving rapidly, Walmart, attracted by the opportunity to continue its expansion, offered $225 million to acquire Cornershop [5]. Cornershop accepted the offer in September 2018 and will be able to take advantage of Walmart´s expertise in machine learning to improve their predictions models and make the operations more efficient. Going forward, Cornershop will have to analyze which is the best way to integrate the data science team of both companies and will have to test the machine learning methods developed by Walmart. In addition, when expanding into new cities, Cornershop will have the opportunity to use both companies´ databases as a starting point for developing their predictions.
Continuing the efficient expansion of Cornershop
In the future, the management of Cornershop should continue focusing on the expansion of the company and on cost reductions. Currently, the users are mostly in a high socioeconomic demographic [6]. Since the use of credit cards in the middle and low-income segments is low, Cornershop should offer other payments methods such as cash or electronic deposits. In addition, these segments will show completely different purchase behaviors, which will have to be used to train the model. To illustrate, the amount and the variety of the purchases will be lower, but the frequency of the purchases could be higher.
With the objective of reducing the costs, Cornershop could use machine learning methods to optimize the purchases of the delivery staff. The company could use past purchase behavior to motivate delivery staff to be in zones with expected high demand, reducing the transportation times. Along the same line, Cornershop could manage the demand and distribute purchases throughout the day by charging different delivery prices to the customers.
The improvement of the prediction models appears to be an infinite iterative process. Which other parts of the process could be improved through machine learning? Which other machine learning methods could Cornershop´s team apply to increase the efficiency of the processes?
(Word count: 797)
[1] Cornershop website, “Nuestra historia,” https://cornershopapp.com/about, accessed November 2018.
[2] Cornershop website, “Mejorando la estimación de tiempo usando Machine Learning,” https://tech.cornershop.io/mejorando-las-estimationes-de-tiempo-parte-1-1af96d8afb4c, accessed November 2018.
[3] Fedyk, “How to tell if machine learning can solve your business problem.” Harvard Business Review Digital Articles (November 15, 2016).
[4] Forbes. “What are Neural Networks – A simple explanation for absolutely anyone,” https://www.forbes.com/sites/bernardmarr/2018/09/24/what-are-artificial-neural-networks-a-simple-explanation-for-absolutely-anyone/#53cffad11245, accessed November 2018.
[5] Forbes, “Why is Walmart Acquiring Cornershop,” https://www.forbes.com/sites/greatspeculations/2018/09/14/why-is-walmart-acquiring-cornershop/#717c6f233e32, accessed November 2018.
[6] Economia y Negocios, “La historia y el modelo tras el despegue de Cornershop,” http://www.economiaynegocios.cl/noticias/noticias.asp?id=495340, accessed November 2018.
Hello Andres, I found your article very interesting. The problem of last mile delivery is a key problem for companies taking on the challenge of delivering same day through independent contractors. Solving the problem is particularly challenging because, as you also pointed out, there is a high number of factors increasing variability in delivery times: the size of the order, how many unique SKUs the customer is choosing, how busy is the supermarket at a particular time of the day, how much traffic there is on the streets, how far is the driver from the supermarket, the cumulative delay coming from the fulfillment of different orders in a row, whether the picking and delivery tasks are done by the same person and many more. The use of neural networks here is certainly beneficial but since you asked about what other techniques could be considered to make the process even more efficient, I would highly recommend Cornershop to keep a close eye on what other peers in the same industry around the world are doing to solve the issue. For example, check out this article wrote by the data science team at Instacart on the usage of quartile regressions to improve delivery times: https://tech.instacart.com/how-instacart-delivers-on-time-using-quantile-regression-2383e2e03edb. This is just one of the many articles Instacart posted on the topic and they are all available online.
This article is insightful! One of the main issues in the retail industry today is what Cornershop has – perhaps – solved. I would want Walmart to dive into the mid and low-income issue you describe, as it is these sectors that likely drive most volume in the market. Most likely the calculations will have to include higher controls of profitability as the volumes will increase disproportionately to revenue when we apply these techniques to lower income buyers.
This is an exciting development and I will make sure to stay informed.