Sales prediction

Performing analysis on Sales Prediction dataset and make prediction using machine learning models.

Author: Kim Hazed Delfino

Business problem:

Uncover some potential KPI's (Key Performance Indicator) between store location, type, item-type and other data collected to accurately predict total sale of an item and ultimately improve revenue by providing data-driven-decision and make "what makes this item/store successful" reproducable for future business iteration.

Methods

• remove all duplicates to avoid repeating data
• correct all data inconsistencies (mislabled, unnecessary extra character)
• Impute missing values to keep data integrity and avoid model prediction errors

Exploratory Data Analysis

• Quick glance of Item MPR Distribution:

Here we can see the normal distribution of item Maximum Retail Price (MRP) and get a better understanding of where majority of our item are priced at

---

This graph represents the average sales per outlet_type and Location_type (Tier). In terms of perfomance it looks like Supermaket Type3 paired with Location_type Tier3 outperforms other types and tiers

---

Here we can see that the top3 ave item_type sales belongs to Starchy Foods, Seafood, and Fruits and Vagetables respectively

---

Models

• Linear Regression - Simple linear regression is a regression model that estimates the relationship between one independent variable and one dependent variable using a straight line.

• DecisionTree Regressor - A decision tree is a type of supervised machine learning used to categorize or make predictions based on how a previous set of questions were answered.

Recommendations:

• more data means more opportunity to find commonality for analysis and paterns for model predictions, we can't go wrong with adding more data in our dataset
• gather more relavant data like monthly sales to determine wheter a certain item is better suited as seasonal item or staple item in the store

Limitations & Next Steps:

• the use of simple machine learning model limits our capability to fine-tune our models to better understand patterns in our dataset
• next Step - Explore different and more complex machine learning models like RandomForest Regressor , Bagging Regressor , Elastic_Netto get a better feel and benchmark different model performance to one another.

---

Model Observation Plots

The 3 largest coefficients are for

• "Outlet_Identifier_OUT027" for every item outlet_027 have, their predicted Item_outlet_sale Increases by $588.235
• "Outlet_Type_Supermarket Type3" items sold in Supermarket Type3 increases the Item_outlet_sale average by $588.235 (similar value to OUT027 because OUT027 is the only Supermarket Type3 in the Dataset)
• "Outlet_Size_Medium" Items sold in Outlet size medium adds Item_Outlet_sale avg by $421.406

What were your 5 most important features?

• The top 5 most important features are:
  • item_MRP - single most important
  • Outlet_Type_Grocery Store - 2nd most important
  • Item_Visibility - somewhat important
  • Item_Weight - somewhat important
  • Outlet_Type_Supermarket Type3 - somewhat important

• As we can see above, SHAP's Top-2 importances are similar with built-in random forest importances but the rest are not the same.

• Reading SHAP Top3 Summary Plots
  • "Item_MRP" - Higher(red) item_MRP have positive impact on our model and opposite can be said to lower(blue) feature value item_MRP
  • "Outlet_Type_Grocery Store" - Items that are sold under outlet_type Grocery store have negative impact on our model output which make sense because based on our EDA Grocery store has the smallest distribution relative to other outlet_type
  • "Outlet_Type_Supermarket Type3" - the opposite of Grocery store, Supermarket Type3 holds the top spot of highest avg item_outlet_sales compared to other outlet_type based on our EDA which explain why it has high possitive impact on our model output

• As we can see in the LIME explanation above, there were many factors contributing to the predicted Item_Outlet_sale, such as:
  • Is the Outlet type - Grocery store (False - positive impact)
  • Item_MRP (low MRP - negative impact)
  • Is the item sold by OUT027 (True - positive impact)

• Here we can see how much each feature affects the model prediction for example Row 39
  • 54.01(min)~7562.66(max)
  • Item_MRP has the most influence driving our prediction to be on the lower-end relative to base value due to having lower Item_MRP value
  • Outlet_Identifier_OUT027 and Outlet_Type_Supermarket Type3 have the 2nd and 3rd most influence with our model prediction respectively, Pushing our model to stay closer to the base value offsetting the negative impact of item_MRP.

• As we can see in the LIME explanation above, there were many factors contributing to the predicted Item_Outlet_sale, such as:
  • Is the Outlet type - Grocery store (True - Negative impact)
  • Item_MRP (low MRP - negative impact)
  • Is the item sold by OUT027 (False - Negative impact)

• Here we can see how much each feature affects the model prediction for example Row 96:
  • 52.98(min)~7575.22(max) / Predicted Value : 157.18
  • Outlet_Type_Grocery Store has the most influence driving our prediction to be on the lower-end relative to base value due to the item being sold at a grocery store and having lower Item_MRP value

---

For further information

for any additional question, please contact me at my email