store-sales-forecast

Predicting store's sales

1. Business Problem

Rossmann is one of the largest drug store chains in Europe with around 56,200 employees and more than 4000 stores across Europe.

Rossmann store managers often have the difficult task of predicting their daily sales up to six weeks in advance. Store sales can be influenced by many factors, including:

• promotions, competition, school and state holidays, seasonality and location,...

Additionally, there are thousands of individual managers predicting sales based on their unique circumstances. As a result, in 2015, Rossmann saw that the accuracy of these prediction results could be quite varied and created a kaggle competition to challenge data scientists to create an efficient model that would help them with this all-important prediction.

2. Business Assumptions

The assumptions about the business problem is as follows:

• "Sales are the lifeblood of a business. It's what helps you pay employees, cover operating expenses, buy more inventory, market new products and attract more investors. Sales forecasting is a crucial part of the financial planning of a business. It's a self-assessment tool that uses past and current sales statistics to intelligently predict future performance."

• "With an accurate sales forecast in hand, you can plan for the future. If your sales forecast says that during December you make 30 percent of your yearly sales, then you need to ramp up manufacturing in September to prepare for the rush. It might also be smart to invest in more seasonal salespeople and start a targeted marketing campaign right after Thanksgiving. One simple sales forecast can inform every other aspect of your business."

• "Sales forecasting is one of the most important business processes to running the business. It determines how the company invests and grows and can have a massive impact on company valuation."

References:

• https://money.howstuffworks.com/sales-forecasting1.htm
• https://www.clari.com/blog/the-importance-of-sales-forecasting/

3. Solution Strategy

1. Business question:
   • Forecast store sales in the next 6 weeks
2. Understanding the business (it was invented to further contextualize our problem):
   • Find out where the root problem comes from → the CFO needs to renovate the stores and he needs to know which ones he should invest the most
     • For this, the forecast of 6 weeks of sales seems reasonable as a solution
3. Data collection
   • Available in kaggle
4. Data cleaning
   • Description of data
     • Knowing the size of the problem we are facing
   • Feature engineering
     • Create new variables from the originals, aiming at a better visualization of them
   • Filtering of variables
     • Filter variables based on business bias (data that cannot be used by the company in general, data that will not be available at the time of prediction,...)
     • This influences how we will model the algorithm
5. Exploratory data analysis( EDA )
   • Understand the business from the point of view of data
   • Create the feeling, through plots and hypotheses, of which variables would be important for learning the model
     • We generate insights, even causing a possible breach of beliefs in other employees
6. Data modeling
   • We separated the data, but now we need to prepare the data for machine learning models
     • Encoding of categorical and numerical variables
   • Feature selection using some ML algorithm, in this project it was Boruta
     • We see the correlation of the analyzed variable with the answer
7. Machine Learning Algorithms
   • We initially implemented 1 baseline algorithm, 2 linear and 2 non-linear
   • After that we apply crossvalidation to validate the performance of the models we choose
8. Algorithm Evaluation
   • We analyze the error from the point of view of model performance
     • Looking at MPE, RMSE, MAE, MAPE

4. Top 3 Data Insights

1. Stores with closer competitors should sell less.(common sense)

   • FALSE, Stores with CLOSER COMPETITORS sell MORE.
     • Is it important for the model? It might be important for the model, but in a weaker way

2. Stores with larger assortments should sell more.

   • FALSE, because stores with promotions active for a long time sell less after a certain period of promotion.
     • Is it important for the model? No, low pearson

3. Stores should sell more over the years.

   • FALSE, Stores sell less over the years.
     • Is it important for the model? Yes, very high correlation

5. Machine Learning Model Applied

Tests were made using different algorithms.

6. Machine Learning Model Performance

The chosen algorithm was the XGBoost Regressor. In addition, I made a performance calibration on it.

MAE, MAPE and RMSE

These are the metrics obtained from the test set.

MAE	MAPE	RMSE
664.974996	0.097529	957.774225

The summary below shows the metrics comparison after running a cross validation score with stratified K-Fold with 10 splits in the full data set.

Model Name	MAE	MAPE	RMSE
Random Forest Regressor	837.68 +/- 219.1	0.12 +/- 0.02	1256.08 +/- 320.36
Linear Regression Regularized Model - Lasso	2117.26 +/- 341.08	0.29 +/- 0.01	3061.0 +/- 503.47

7. Business Results

You can see the results of the forecasts on your own phone (or computer) by accessing the bot on the telegram that will make the forecast in real time

• Link to access the results=> t.me/rossmann_pred_edu_bot
• type /number_of_store (1 through 1115) to receive the sales prediction for that store

10. Next Steps to Improve

1. Use more models to lower error metrics .