• Acquiring thorough profit insights is critical in the current competitive online shopping platform scenario. These insights are crucial for developing successful sales tactics and offers, which in turn lead to higher sales and more profits.

• We built a set of functions using linear regression, Elastic Net model, neural network model and Random Forest model on top of Amazon sales data. With the use of these features, we can identify weekly profit trends depending on consumer actions. Our ability to accurately and precisely improve sales performance can be enhanced by utilizing this analytical method.

Question: Can we develop a profit function using the provided sales data that offers insights into profitability, considering various aspects such as purchase day, repeat purchases, and product categories?

Four distinct models were utilized in the search for the best function, hence a large number of libraries were required.

suppressWarnings({
library(ggplot2)
library(dplyr)
library(tidyr)
library(readxl)
library(nnet)
library(modelr)
library(tidyverse)
library(glmnet)
library(randomForest)
library(knitr)
library(caret)
library(kableExtra)
})

We used read_excel to import the data and altered the column names for more practical purposes..

sales.data <- read_excel("Amazon_2_Raw.xlsx")
colnames(sales.data) <- c("order_id","order_date","ship_date","email","geography","category","product_name",
                          "sales","quantity","profit")

It is necessary to organize the data in the right format because there are numerous figures and information that were separated.

sales.data$order_date <- as.Date(sales.data$order_date)
sales.data$ship_date <- as.Date(sales.data$ship_date)
sales.data$year <- format(sales.data$order_date, "%Y")
sales.data$month <- format(sales.data$order_date, "%m")
sales.data$day_of_week <- weekdays(sales.data$order_date)
sales.data$shipping_duration <- as.numeric(difftime(sales.data$ship_date, sales.data$order_date,
                                                    units = "days"))
sales.data$year <- as.factor(sales.data$year)
sales.data$month <- as.factor(sales.data$month)
sales.data$day_of_week <- as.factor(sales.data$day_of_week)
sales.data$category <- as.factor(sales.data$category)
sales.data$sales_normalized <- scale(sales.data$sales)
sales.data$quantity_normalized <- scale(sales.data$quantity)
sales.data$shipping_duration_normalized <- scale(sales.data$shipping_duration)
sales.data$order_day <- weekdays(as.Date(sales.data$order_date))
sales.data$repeat_purchases <- ave(sales.data$order_id, sales.data$email, FUN = length)
sales.data$repeat_purchases <- as.numeric(as.character(sales.data$repeat_purchases))
sales.data$repeat_purchases_normalized <- scale(sales.data$repeat_purchases)
sales.data$State <- sapply(strsplit(as.character(sales.data$geography), ","), function(x) x[3])

We want to analyze customer purchasing behavior over time, specifically identifying peak periods of purchases and lows throughout the year. I utilized a line graph plot for easy visualization of the peaks and troughs in customer purchasing patterns.

• Varied Sales Volume: The bar chart demonstrates significant variation in total sales among categories, with ‘Chairs’ and ‘Phones’ leading in sales volume.
• Focus Areas for Growth: Categories such as ‘Fasteners’, ‘Envelopes’, and ‘Labels’ show lower sales, potentially indicating areas for strategic sales initiatives or product development.

• Profit Leaders: ‘Copiers’ and ‘Accessories’ categories show the highest total profit, indicating strong performance and possibly higher margins.
• Areas for Improvement: ‘Bookcases’ and ‘Machines’ represent the lowest profit categories, suggesting a need for a review of pricing strategies, cost control, or sales tactics.

Long term sales have increased but short-term behavior has been volatile. Profit, however, has been steady with a spike in early 2014. This data shows how Amazon has performed over time and can be used to predict its future value.

After geographical analysis, California orders the most Amazon products by far. Washington is second, with the other states ordering similar amounts. This data could help Amazon build more facilities in California, where product demand is higher.

The dual-axis plot above shows sales and profit by product category, with bars representing sales and lines with markers representing profit. This visualization shows how sales translate into profit across product categories. From the plot, we can see each category’s sales and profit, which may help us decide which product categories to focus on or expand.

• The linear regression model uses category, repeat_purchases, and quantity as predictors to estimate profit.
• The Mean Absolute Error (MAE) is 47.39, indicating that the model’s predictions are, on average, $47.39 away from the actual profit values.
• The Mean Squared Error (MSE) is reported at 25325.26, suggesting that the model’s predictions are quite variable and may be influenced by outliers or extreme values.
• The R-squared value of the model is 0.1643, which means that about 16.43% of the variation in profit is explained by the model. This is a relatively low value, implying that the model might not be capturing all the factors that influence profit.

• Pattern in residuals: The spread of residuals increases with fitted values, suggesting non-constant variance.
• Presence of outliers: Specific data points, such as 2664, stand out with high residuals.

• Distribution issues: The tail of the plot curves upwards, thus not a normal distribution.
• Influential observations: Points like 2664 stray from the expected line, possibly exerting undue influence on the model.

Neural Network Training Introduction

• Goal: Predict target using a neural network model.
• Data: Features prepared from sales.data.
• Method: 5-fold cross-validation to evaluate model performance.

Data Preparation and Model Configuration

• Data matrix: data_for_nn prepared with predictors.
• Target variable: profit from sales.data.
• Neural network setup: nnet with size=5, decay=0.000005, max iterations=3000.

Training Output Summary for Each Fold

• Fold 1: Initial value: 44,034,439.54 | Final value: 35,729,033.63
• Fold 2: Initial value: 88,105,695.11 | Final value: 58,442,771.49
• Fold 3: Initial value: 93,857,949.85 | Final value: 71,036,755.92
• Fold 4: Initial value: 94,524,748.27 | Final value: 74,115,030.54
• Fold 5: Initial value: 82,663,911.40 | Final value: 56,772,514.61

Average Performance Across All Folds

• Average Mean Absolute Error (MAE) across all folds: 50.21544
• Average Mean Squared Error (MSE) across all folds: 31,663.92

Elastic Net Model Overview

• Hybrid Model: Combines Lasso and Ridge attributes with alpha at 0.5.
• Average Error: MAE of $39.38 shows the model’s close-fit predictions.

Model Performance

• Precision: An MSE of 13,956.61 reflects a robust predictive accuracy.
• Explained Variance: R-squared value at 53.94%, a substantial fit indicating the model’s effectiveness.

Prediction Alignment: “Actual vs. Predicted” visuals show many predictions closely match actual profits.

Residual Trends: “Residual Plot” reveals patterns suggesting the influence of outliers on prediction accuracy.

• High Variance Explained: With an R-squared of 85.5%, the model accounts for a significant portion of profit variance.
• Model Complexity: Utilizes 500 trees to achieve strong predictive accuracy and fit.

## Random Forest - Mean Absolute Error (MAE): 13.73329

## Random Forest - Mean Squared Error (MSE): 4395.276

## R-squared: 0.854963

The “Actual vs. Predicted Profits (Random Forest)” plot illustrates a tight clustering of points along the red line that indicates perfect predictions, demonstrating the model’s accuracy in predicting profits.

In the “Residual Plot (Random Forest)”, most residuals are distributed close to the horizontal line at 0, with fewer large residuals compared to previous models, suggesting that the Random Forest model has a consistent prediction quality across the range of actual profits.

• Leverage Predictive Insights: Utilize the Random Forest model’s feature importance to prioritize variables like category, quantity, and sales_normalized that are most predictive of profit.

• Optimize Product and Sales Mix: Focus on high-profit categories and optimize inventory levels using insights from sales_normalized and quantity_normalized to meet demand without overstocking.

• Refine Temporal Strategies: Analyze patterns from year, month, and day_of_week to capitalize on peak shopping times and tailor marketing efforts.

• Efficient Logistics: Streamline the shipping process based on shipping_duration_normalized to improve customer satisfaction and reduce costs, thereby enhancing profit margins.