Retail Sales Dataset Project

This project is part of the Data Science Certificate program at the University of Toronto’s Data Sciences Institute.

Dataset

We are using the Retail Sales Dataset, which contains transactional-level sales data including information such as the date of purchase, customer gender and age, the type of product bought (product category), how many items were purchased, the price per item, and the total revenue generated from each transaction. This data allows us to explore shopping patterns and customer behavior across different groups.

The raw dataset is stored in the data/raw/ directory.

Team Members

• Iryna Verbova (iverbova)
• Tetiana Hakansson (t125yf)
• Rabia Imra Kirli Ozis (rabiaimra)
• Ting Man (manntintin)
• Vrushali Patil (vrushalipatil9763)

Business Problem

This project investigates purchasing behavior in the context of a small to medium-sized retail business. We aim to explore how customer demographics (age and gender) and product features (category and price) affect purchasing patterns.

Specifically, we are exploring:

• How sensitive different age groups and genders are to price changes.
• Whether it's possible to identify high-spending customers based on transaction attributes.

The outcomes will help retail stakeholders make more informed pricing decisions, segment customers, and design targeted marketing strategies.

Stakeholders

The primary stakeholders who will benefit from this project include:

• Retail business owners and managers looking to better understand customer behavior and optimize sales performance
• Marketing teams designing promotions and loyalty programs
• Pricing analysts interested in understanding how different customer groups respond to price changes

These stakeholders are focused on improving profitability, increasing customer retention, and delivering more targeted, data-driven customer experiences.

---

Business Questions

1. Regression Question: Price Sensitivity by Customer Segment

   Can we predict whether price per unit affects the quantity purchased differently across age groups or genders?
   → This helps us identify which customer segments are more price sensitive, supporting smarter pricing decisions.

2. Classification Question: Identifying High-Spending Customers

   Can we classify whether a customer will be a high spender based on their demographics and purchase details?
   → This enables targeted marketing and personalized engagement with valuable customers.

---

Initial Dataset Review

• The dataset is clean, well-structured, and ready for analysis.
• No missing values or duplicates.
• Total Amount is consistent with Quantity × Price per Unit.
• Categorical fields (Gender, Product Category) are well-defined and usable.
• Contains sufficient diversity in ages, product categories, and purchase amounts, making it suitable for exploring both regression and classification questions.

Methods & Technologies

• Version Control & Collaboration: Git, GitHub (branching, PR review workflow).
• Data Collection & Storage: CSV datasets organized into raw and processed data folders.
• Data Exploration & Visualization: Python (Pandas, NumPy, Matplotlib, Seaborn), correlation heatmaps, bar and pie charts, time-series plots.
• Feature Engineering: Categorical encoding, feature creation, interaction terms, and scaling.
• Statistical Modeling & ML:
  • Regression: Ordinary Least Squares (OLS) with diagnostics (residual analysis, R², p-values).
  • Classification: K-Nearest Neighbors (KNN), confusion matrix, error rate vs. k-value analysis.
• Evaluation Metrics: R², residual analysis, confusion matrix, error rate.
• Notebook Tools: Jupyter Notebooks for data exploration, feature engineering, regression, and classification workflows.
• Documentation: Markdown, structured project folders (/data, /images, /notebooks).

Risks and Uncertainties

• The dataset may lack granularity for some modeling tasks.
• The sample size is relatively small (1,000 rows), which may affect the accuracy and reliability of our classification model.
• A small sample size in a regression model can lead to unreliable coefficient estimates, inflated standard errors, and reduced statistical power, making it difficult to detect true relationships between variables.
• Since the dataset used in this project is implied to be synthetic, it may not fully capture the complexity, variability, and noise present in real-world data, which could limit the generalizability of our model's performance.
• There is a risk of overfitting if too many derived features are included without careful validation.
• Relationships between variables (e.g., age, gender, and quantity purchased) may be weak or non-linear, which can reduce model effectiveness.

---

Project Tasks & Responsibilities

Team Member	Task
Iryna	Data exploration, feature engineering
Imra	Conducting regression modeling
Vrushali	Conducting classification modeling
Tetiana	Writing conclusions, maintaining narrative
Mandy	Presentation development

Our project timeline proceeded as planned, with all tasks completed according to schedule.

Review Process

To ensure high-quality collaboration, we followed a structured pull request (PR) review process:

• Each member created PRs for their work (data exploration, feature engineering, modeling, etc.).
• Reviewers were assigned based on which parts of the project were impacted:
  • If a PR impacted only one or two people’s work, those members were assigned as reviewers.
  • For PRs affecting the entire project (e.g., narrative updates or final clean-up), the entire team was assigned as reviewers.
• A PR could only be merged once all assigned reviewers approved it.
• The creator of the PR resolved any merge conflicts and performed the final merge.

This approach ensured accountability, improved code quality, and allowed all team members to be actively involved in the review process.

---

Data Exploration

We take an initial look at the Retail Sales Dataset to uncover patterns and trends and provide valuable insights that will guide future analysis.

Objectives:

• Examine the structure and contents of the dataset
• Create exploratory visualizations to reveal patterns in purchasing behavior
• Produce statistical summaries to identify key trends

Hence, our main goal is to explore how factors like product price, product type, age, and gender influence customer buying habits and total sales performance.

Dataset Overview

We worked with a clean dataset of 1,000 retail transactions. The key features included product information, price, quantity, and demographic data like age and gender.

• Transaction ID: Unique identifier per transaction
• Date: Transaction date
• Customer ID: Anonymized customer identifier
• Gender: Customer gender
• Age: Customer age
• Product Category: Product type (Beauty, Clothing, Electronics)
• Quantity: Number of items purchased
• Price per Unit: Cost per item
• Total Amount: Transaction value = Quantity × Price per Unit
• All columns are well-formatted. There are no missing or duplicate values, and the calculated field Total Amount is consistent with Quantity * Price per Unit.

Exploratory Data Analysis (EDA)

We conducted extensive exploratory data analysis (EDA), including:

• Sales Trends: Time-series analysis of total sales over the year.
• Customer Segmentation: Sales by product category, gender, and age group.
• Key Metrics: Average order value (AOV), average price per unit, and correlation between numerical features.

Exploratory Insights:

Understanding these patterns can help businesses adjust pricing, better segment customers, and tailor campaigns more effectively.

Key Visualizations

Sales Trend: Sales fluctuate across dates with noticeable spikes around mid-year:

Product Category: Electronics contributes the highest revenue, followed by Clothing and Beauty:

Gender Split: Females account for slightly higher overall sales compared to males:

Age Groups: Customers aged 40-60 make the largest contribution, followed by those aged 25-40, who also significantly contribute to sales volume:

Average Order Value (AOV): $456.00.

Price per Unit: Beauty items have the highest average price per unit, while Clothing is the most affordable category.

These findings will guide our feature engineering and modeling decisions in the next phase.

---

Feature Engineering

Turning Raw Data into Insightful Features

To help our models make better predictions, we first needed to enhance the dataset by creating new, more informative features — this process is called feature engineering. We transformed the original retail data into features that could be used in further analysis. These new features help our models recognize patterns more clearly, such as who is more likely to make large purchases or when sales tend to peak.

-We will generate all relevant features needed for our classification and regression models, including:

• Categorical encodings
• Temporal features
• Interaction-ready variables
• Target variable for classification

At the end, we’ll export a clean processed_data.csv to be used for modeling.

To simplify linear regression modeling, we also create numeric encodings for:

• Gender: Male = 0, Female = 1
• Age Group: <25 = 1, 25-40 = 2, 40-60 = 3, 60+ = 4
• Product Category: Clothing = 1, Electronics = 2, Beauty = 3

Feature	Type	Reason
Age Group	Categorical	Needed to evaluate interaction effects between age and price in regression. Binned into <25, 25-40, 40-60, 60+ for business relevance.
High Spender	Binary	Target variable for classification. Labeled as 1 if the transaction is in the top 25% of Total Amount.
Month	Numeric	Temporal feature to explore monthly patterns or seasonality.
Day of Week	Numeric	Helps identify weekday/weekend trends. Can be used to enrich predictions.
Avg Price per Item	Numeric	Provides insight into pricing behavior per transaction.
Gender_, ProductCategory_, AgeGroup_*	One-hot encoded categorical vars	Useful for classification models and non-linear ML algorithms. drop_first=True used to avoid dummy variable trap.
Gender_Num, AgeGroup_Num, ProductCategory_Num	Numeric (label encoded)	Added to support regression models (linear models often benefit from single numeric representations of categories).

Example:

From original:	After Transformation:
Age = 43	Age Group = 40–60
Gender = Female	Gender = 1
Product = Electronics	Product Code = 2

This is what feature engineering looks like — we convert human-friendly info into model-friendly numbers.

Our goal is to teach the model to recognize which customers are likely to spend more — so we created this target feature. With our engineered dataset, we can move confidently into building models that provide useful insights for decision-making.

---

Regression Model

To better understand the factors that drive product purchase quantity in a retail context, we developed a linear regression model using transactional data. The goal was to examine how unit price, customer demographics (age group and gender), and their interactions influence the quantity of items purchased in a single transaction.

Model Specification:
Quantity = β0 + β1*(Price per Unit) + β2*(Age Group) + β3*(Gender) + β4*(Price per UnitAge Group) + β5*(Price per Unit*Gender) + ε

This specification allows us to capture both the main effects of price, age, and gender, and how these effects interact — particularly, whether price sensitivity differs by demographic group.

The interaction terms capture differential price sensitivity, such as how price impacts might vary between age groups or between male and female customers. This allows the model to go beyond average effects and identify segment-specific behavioral patterns, which are often crucial for effective targeting and personalization in modern retail strategies.

From an industry perspective, this type of model provides valuable insights into price sensitivity across different customer segments. For example:

• Personalized pricing and promotions: Marketing teams can use insights from interaction terms to identify which segments (e.g., age 25–40) are more price-sensitive and target them with dynamic pricing or customized discounts.
• Demand forecasting: Understanding which demographic groups are more responsive to price helps inventory planners and analysts model future sales volumes more accurately under different pricing strategies.
• Segmentation analysis: The model provides evidence-based support for customer segmentation strategies by quantifying behavioral differences in purchase patterns across age and gender groups.
• Strategic decision-making: Retailers can use such models to inform store-level pricing, online targeting, and cross-channel optimization efforts based on predicted consumer responsiveness.

While the model’s predictive power is limited, it offers a foundational framework for interpreting customer behavior and can be expanded by incorporating additional variables such as seasonality, product-level features, loyalty data, or promotional history. In real-world applications, such models could be integrated into business intelligence dashboards or dynamic pricing systems to support evidence-based decision-making.

OLS Regression Results

Metric	Value	Metric	Value
Dep. Variable	Quantity	R-squared	0.014
Model	OLS	Adj. R-squared	0.002
Method	Least Squares	F-statistic	1.172
Date	Fri, 18 Jul 2025	Prob (F-statistic)	0.310
Time	19:39:29	Log-Likelihood	-1148.1
No. Observations	750	AIC	2316
Df Residuals	740	BIC	2362
Df Model	9	Covariance Type	nonrobust

---

Coefficients Table

Variable	coef	std err	t	P>|t|	[0.025	0.975]
const	2.1543	0.163	13.184	0.000	1.834	2.475
Price per Unit	0.0017	0.001	2.715	0.007	0.000	0.003
Gender_Num	0.0658	0.113	0.584	0.559	-0.155	0.287
Age Group_25-40	0.3823	0.188	2.038	0.042	0.014	0.750
Age Group_40-60	0.3639	0.176	2.069	0.039	0.019	0.709
Age Group_60+	0.1678	0.215	0.781	0.435	-0.254	0.590
Price_Age_25_40	-0.0016	0.001	-2.285	0.023	-0.003	-0.000
Price_Age_40_60	-0.0019	0.001	-2.844	0.005	-0.003	-0.001
Price_Age_60+	-0.0010	0.001	-1.146	0.252	-0.003	0.001
Price_Gender_Interaction	-0.0003	0.000	-0.617	0.537	-0.001	0.001

---

Model Diagnostics

Test	Value	Test	Value
Omnibus	6041.210	Durbin-Watson	1.949
Prob(Omnibus)	0.000	Jarque-Bera (JB)	57.838
Skew	0.023	Prob(JB)	2.76e-13
Kurtosis	1.640	Cond. No.	2.63e+03

Notes: [1] Standard Errors assume that the covariance matrix of the errors is correctly specified. [2] The condition number is large, 2.63e+03. This might indicate that there are strong multicollinearity or other numerical problems.

Model Objective:

Predict Quantity purchased based on unit price, age group, gender, and their interactions.

Model Fit:

• R² = 0.014, Adj. R² = 0.002 → The model has weak explanatory power.
• F-statistic: 1.172, p = 0.310 → Our predictors jointly do not explain the variation in quantity.
• Residual analysis indicates potential model misspecification.

Key Findings:

• Price per Unit: Higher prices slightly increases quantity on average. The effect is small but statistically significant (p = 0.007).
• Gender (Female vs. Male): No significant effect of gender-based differences on quantity purchased.
• Age Groups: Customers aged 25–60 purchase slightly more than those under 25.

Interaction Terms:

• Higher prices reduce quantity more for age groups 25–60.
• No evidence of gender-based differences in price sensitivity.

Interpretation for Business & Takeaways

• Some individual coefficients, such as price per unit, age groups 25-40 and 40-60, and their interaction terms, are statistically significant which implies that they have significant explanatory power in explaining the quantity purchased.
• However, the overall model lacks predictive power with low R2 and F-statistic. Thus, our model is not explaining the variability in quantity well. Most of the variation is likely driven by factors not included in the model.
• Interaction terms reveal age-based differences in price sensitivity.
• Gender appears to have no meaningful effect on quantity purchased.
• Interestingly, the regression model reports a positive and statistically significant main effect of price per unit, suggesting that - all else equal - higher-priced products are associated with larger quantities purchased. While this may seem counterintuitive from a classical demand perspective, it could reflect nonelastic or luxury-oriented purchasing behavior, or the presence of premium product categories where higher prices do not deter demand and may even be associated with perceived value.
• However, the model also includes interaction terms between price and age group, and these show significant negative coefficients for customers aged 25–60. This indicates that for these segments, higher prices are associated with lower quantities purchased, aligning more closely with traditional downward-sloping demand behavior. In other words, price sensitivity does exist in the data, but it varies across customer segments.

Additional Notes

• To assess the model's ability to generalize beyond the data it was trained on, we randomly split the dataset into a training set (75%) and a test set (25%). The results indicate that the model performs poorly on unseen data with considerable percentage error in predictions.

---

Classification Model

To identify high-spending customers, we built a K-Nearest Neighbors (KNN) classifier.
The aim was to determine whether we can predict if a customer is likely to be a high spender based on their demographics (age, gender) and purchase attributes (product category, quantity, and price).

---

Approach:

Defining High Spenders:
We labeled the top 25% of transactions by total spending as "High Spenders" (1), while the rest were labeled "Not High Spenders" (0).

Feature Engineering:
The model used our processed dataset with:

• Encoded categorical variables (Gender, Product Category, Age Group).
• Numerical features (Quantity, Price per Unit).
• A binary target variable (High Spender).

Data Split:
We split the dataset into 75% for training and 25% for testing to evaluate how well the model generalizes to new data.

Standardization:
Because KNN compares data points based on distance, all features were standardized to have comparable scales. For example, age and quantity were scaled to ensure that no single feature dominates the distance calculation.

Selecting K (Number of Neighbors):
We tested K values from 1 to 20 and plotted the error rate vs K to find the optimal balance between overfitting (K too small) and underfitting (K too large).
The optimal K = 2 was chosen for best performance.

The best value of k is: 2

Train KNN We create a KNN classifier using k neighbors after finding the best value for K by Hyperparameter Tuning. KNN works by comparing each test point to the k closest training points and assigning the majority class among them.

KNN Distance Calculations: Compute distance to all training points using Euclidean distance formula: distance: $d = \sqrt{(x_1 - x_1')^2 + (x_2 - x_2')^2 + \dots + (x_n - x_n')^2}$

It picks K closest points from training set using the above distance. For classification, KNN checks the most frequent class among the k nearest neighbors.

Example:

If k=5 and the nearest neighbors have labels [1, 0, 1, 1, 0] → the predicted label is 1 (majority). In case of a tie, behavior depends on the implementation (e.g., some libraries break ties by choosing the class with the lower label).

Confusion Matrix

Confusion Matrix:
This heatmap shows how well the classifier predicted the two classes:

• True Positives (TP): 11 (correctly identified high spenders).
• True Negatives (TN): 191 (correctly identified non-high spenders).
• False Positives (FP): 9 (non-high spenders incorrectly predicted as high).
• False Negatives (FN): 39 (missed actual high spenders).

It helps identify if the model is biased toward one class or struggles with imbalanced data.

Model Performance Summary

Metric	Value
Best k	2
Accuracy	80.8%
Precision (High Spenders - Class 1)	55%
Recall (High Spenders - Class 1)	22%
F1 Score (High Spenders)	31%

Results:

• Accuracy: 80.8% (model correctly predicts 4 out of 5 transactions).
• Precision (High Spenders): 55% (when the model predicts "High Spender," it's correct about half the time).
• Recall (High Spenders): 22% (the model misses many true high spenders due to class imbalance).
• F1 Score: 31% (harmonic mean of precision and recall).

Interpretation for Business & Takeaways

• The model is good at predicting who is not a high spender, which is valuable for excluding customers from premium marketing campaigns.

• However, it misses many true high spenders (low recall), which means we could refine our approach by:

  • Using balanced datasets (e.g., SMOTE for synthetic high spender examples).
  • Trying other algorithms like logistic regression or random forests.

• While this KNN model is simple yet effective, it shows how retailers could identify high-value customers for targeted campaigns. With a larger and more realistic dataset, this model could evolve into a valuable tool for personalized marketing and customer segmentation.

---

---

---

Summary of Findings and Strategic Implications

This project examined how customer demographics (age, gender) and product features (price, product category) influence retail purchasing behavior. We explored spending patterns, engineered features, and built predictive models to answer two key business questions:

1. How does price sensitivity vary across customer segments?
2. Can we identify high-spending customers for targeted marketing?

---

Key Findings

• Data Exploration: Electronics generated the highest revenue, women spent slightly more than men, and customers aged 40–60 were the most active shoppers. These insights provide valuable input for customer segmentation and marketing strategies.

• Feature Engineering: We enhanced the dataset by grouping ages, creating a high-spender label, and encoding categorical variables. These transformations allowed us to prepare the data for regression and classification modeling.

• Regression Model:

  • Price per unit, age groups 25–60, and their interaction terms were statistically significant in explaining purchase quantity.
  • The model’s low R² (1.4%) and F-statistic suggest limited overall predictive power — likely due to factors not present in the dataset (e.g., promotions, seasonal trends).

• Classification Model:

  • Using K-Nearest Neighbors, we achieved 80.8% accuracy, but recall for high spenders was low (22%) due to class imbalance (more low spenders than high).
  • This shows the model is good at identifying non-high spenders but struggles to detect all high-value customers.

---

Business Implications

Although our dataset is implied to be synthetic, the project demonstrates how data-driven approaches can guide real-world retail strategies:

• Pricing Teams can use price sensitivity insights by age group to adjust promotions or discounts.
• Marketing Teams can design campaigns targeting the 25–60 age group and other high-value segments.
• Customer Analytics Teams can adapt similar models to real transaction data for predicting spending behavior and improving customer lifetime value.

---

Future Enhancements

With larger, real-world datasets, these models could be expanded by incorporating:

• Promotional and seasonal trends for better prediction accuracy.
• Advanced classification techniques (e.g., logistic regression, random forests) to better identify high spenders.
• Integration into business intelligence dashboards for ongoing monitoring and strategic decision-making.

---

Takeaway

This project demonstrates a complete data science workflow — from data exploration and feature engineering to modeling and interpretation — and highlights how actionable insights can drive better retail decision-making, even with modest datasets. With real-world data and ongoing refinement, these models could become part of a broader decision-support system, helping retailers make smarter, more targeted choices that improve both customer experience and business sustainability & outcomes.

---

Team Video Reflections

• Iryna Verbova (iverbova)
• Tetiana Hakansson (t125yf)
• Rabia Imra Kirli Ozis (rabiaimra)
• Ting Man (manntintin)
• Vrushali Patil (vrushalipatil9763)