MushroomParachute

Inspiration

We aimed to provide insights for retailers from public datasets to help them with their business. More specifically, we would like to find the importance of different features on the weekly sales of the business, and make prediction about future weekly sales based on those important features.

What it does

We found out several features that could potentially affect the weekly sales of a retailer, they are the following: temperature of the week, price of the fuel, which month or week it is in a year, the CPI of that week, whether it is a holiday or not, whether there was a markdown or not.

The ML models showed the relative importance of different features on the weekly sales of Walmart stores and departments. Based on this information, retailers can understand better how these features affect the weekly sale, so they can act correspondingly to the change in these features. We also used ML models to help retailers make predictions about future weekly sales based on those features.

How we built it

We first found a dataset containing these features and the weekly sales of 45 Walmart stores with different departments. In order to eliminate the effects due to the differences between stores and departments, we chose a specific department in a specific store as our primary working dataset.

We then chose several machine learning models to train this small dataset and obtained the relative importance of the effects caused by the features on the weekly sale of the department in that store. Finally we used a neural network model to make predictions about future weekly sales based on those important features.

Challenges we ran into

We were struggling to find good datasets from the databases, then we searched multiple times in kaggle to find an adequate one. We also were not able to eliminate the impacts of departments and stores on the weekly sales because normally these two would have the most significant impacts, but we extracted smaller data sets trying to solve this problem. However, this led to a lack of data and risks of overfitting since there were multiple features and not enough data to train the model on.

Accomplishments that we're proud of

As emerging participants who were using ML for the first time, we successfully ran several models on the dataset we found and obtained meaningful results.

What we learned

How to clean a dataset before using any models, how to train an ML model on a dataset, what is the general process of training and testing datasets, how to interpret the results from the model that we trained datasets on.

What's next for MushroomParachute

We aim to test the model on more datasets, both in different departments and different stores. We would also incorporate features of the stores and departments and predict their effects on the weekly sale. Also, we would find datasets of other stores to confirm whether the important features we found also had profound effects on other stores.