Sales-Forecast: A Jupyter Notebook repository from mattmagrin

Sales Forecasting the upcoming six weeks

If you wish, you may interact with the final solutions data products I built, just click on the thumbnail below :

--> Created to provide easy and remote access to forecasts.

Project Overview 🎯

Understanding the root cause behind the solution: The company wants to expand the operation, but is not sure how much to invest in each store. The investment needs to be consistent with the projected revenue in each case.

Solution developed: To solve this business problem, I used an adaptation of the CRISP-DM methodology, from which I was able to increase understanding of the influence of factors such as promotion, holiday and competition on revenue. Furthermore, we obtain an assertive prediction in the sales for each storefor the upcoming six weeks.

How was this solution developed? ✔

Each step has a detailed notebook with explanations, just click on the links below

1. Data Description - In this step, after obtaining the data, i did an initial analysis of the data to familiarize myself with the variables involved and to have a clearer view of how to address the problem.

2. Feature Engineering & Data Filtering - Here I tried to create/remove some variables that might help on the understanding and usage of the data, as well as variables that might help M.L. algorithms learn patterns in the data to make predictions.

3. Exploratory Data Analysis (EDA) - Extremely important step to increase my understanding of the business dynamics, formulating and validating hypotheses.

4. Data Preparation & Feature Selection - - In order to obtain more accurate insights, i made some transformations as encodings for categorical data and scaling to numerical ones. Here we have an important trade off, because on the one hand with the rubust scaler we don't depend on outliers, but we have "tighter" distributions. While with the min max scaler we still take into account the outlier but we run the risk of changing the nature of the feature. In addition, I used the Boruta algorithm to help in the initial choice of the features used.

5. Machine Learning Modelling & Hyperparameter Fine Tunning - In this step I used some linear algorithms as well as some tree-based ones, to verify if the phenomena could be modeled linearly or not. In addition, Bayesian optimization was used to find the best hyperparameters for the model with the best performance for the chosen metrics.

6. Business Results & Deploy - In the final step, I adapted the results to the business context and deployed them to the streamlit. We also created the rossman class, to synthesize the transformations, cleanings in the data received via API later. Finally, from the API created, it was possible to launch a bot on telegram to return forecasts, enabling greater accessibility to the solution.

Key lessons learned 🤓

As important as the step-by-step solution to the problem, is the way in which the disposition will be conveyed in a data product / solution.
Models must be chosen while keeping in mind the deployment costs. Sometimes the performance difference won't make up for the risk/costs associated with more development time.
Exploratory data analysis provides not only important information about data and insights to the business team but also an understanding of how some features might help or not the chosen ML models.
Experimenting with the code may give insights on doing things differently, but it must not take forever. Remember Done is better than perfect.

Special thanks