How to predict sales and quanitfy results using machine learning .

We have been hired by a Wal-Mart competitor to determine the best sales strategy and pricing for in-store tolietry projects. So by running the data provided to us through machine learning (ARIMA, LTSM, SARIMA, SVR) we will be able to provide an answer on what items sell best win, what items are most effected by sales promotions and the best way to allocate resources.

Data Origin

• Data provided by an E-commerce company.

Data Included

• Data compiled from 5 Major Retailers in the United States.

Cleaning Data

• Data condensed into 9 categories
• Acne, Bath, Body Wash, Hair, Lotion, Makeup, Shaving, Sunscreen, Tanner
• Removed 2021 Sales Data as we did not have corresponding promotional data to use with it.
• Combined Sales Data and Promotional Data into one .csv file. e
• Further Data Cleaning such as changing column names and transposing was accomplished in our program.

Finalised Data: Sales & Promotional Data 2019/20

• Decompose the data in order to detect the trends in it. See if it is additive or mutiplicative.
• Used Three Main Types of Machine Learning
  • ARIMA
  • LTSM
  • SVR
• ARIMA is used to see the "change since last time." Since it has a moving average the predictions for it will run a bit smoother and likely closely match the actual sales.
• LTSM is the "longest" of our machine learning that we used. The one that sees the biggest picture. While ARIMA takes a snapshot we used LTSM to give us an idea of any greater trends there might be.
• SVR is used to try to establish if there is any link between a specific week of the year and sales in any of our three ideas.
• In our models we kept to 6 as our window and ran 10 Epochs when building the LSTM models. After experimentation we felt this gave us the best results.

• The Sales graph here shows the breadth of our data. All three of the graphs show the same basic pattern of dropping massively during the lockdown, bouncing back up higher then settling down.

• Lotion is our best looking example of a trend in scatter, with it we can see why it has greater than 80% in Seasonality and is our best looking of the SVR predictor.

• As seen here, Arima predicted sales follows the actual sales pretty closely.

• Lotion is the best example of how promotions work. The bulk of the red "no-promo" is on the left side of the graph, smaller sales and the blue "promo" area populates the right side of the graph.

• We rain SVR to see if there was any relation between the sales and any specific week. There did not appear to be any relation to specific weeks on the sales

• All three of the LTSM models predicted that the sales would be consistetenly higher than what it was, likely due to the normal data from 2019/early 2020 and pre-Covid and the massive drop in March 2020 being so outside the normal realm that LTSM did not want to fully commit to the new info.

• Place to put Dashboard/GIFs whenever we design it.

• The Code worked to our expectations. While some data had to be cleaned properly to make sure that the Machine Learning" parts were fed good data everything held up and all the different types of machine learning we ran the info through painted the same general picture.
• We believe that it would be easy to scale the program up to a larger size, or plug and play with any sort of information that you have.

• All three of the items that we closely studied were affected by the COVID-19 Lockdowns of 2020.
• The drastic drop in sales during the COVID-19 lockdown did affect our models, as we can see with the LSTM the predicted sales were consistently above the actual sales, but both sales did follow closely on the same trend lines.
• ARIMA model followed the actual sales the closest in the prediction mode.
• There is no seasonality with Acne or Makeup products, there is slight seasonality with Lotion.
• Promotions were the most effective for Lotion and not very effective for Makeup.
• Even with Lotion having more seasonality than the other you still cannot know for certain what level of sales there will be in a specific week of the year.

• William Brooks
• Jon Iadonisi
• Benjamin McCright
• David Ready