Applied analytics case studies focused on data cleaning, statistical testing, and decision insights, analysing how eBay seller ratings influence customer purchasing behaviour using an inferential (regression) approach.
Key Visual Evidence (click to expand)
1) Missing value assessment (data quality check)
2) Exponential relationship observed in the cleaned data
3) Desmos exponential regression validation (R² ≈ 0.9913)
This applied analytics case study demonstrates how structured data cleaning, statistical modelling, and exponential regression analysis were used to transform a 19,230-row eBay dataset into actionable strategic insight.
After reducing the dataset to 11,018 clean observations, an exponential model achieved an R² ≈ 0.9913, showing a strong relationship between seller rating and review volume (proxy for purchasing behaviour).
The findings support a strategic marketplace decision: small improvements in seller rating near high-performance levels can produce disproportionately large gains in customer engagement and trust.
Business question:
Do existing seller ratings influence other customers’ decision to purchase products?
Hypotheses:
- H0: Existing seller ratings have no impact on customers’ decision to purchase products.
- H1: Existing seller ratings impact customers’ decision to purchase products.
Outcome (summary):
The analysis supports H1, showing a strong exponential relationship between seller rating and review volume (used as a proxy for purchase activity). The insight supports a strategic decision to prioritise rating improvement initiatives to drive customer trust, visibility, and sales performance.
- /report → The full academic case study report (PDF)
- /docs → Extracted visuals and supporting images referenced in the report
Recommended structure:
├── README.md
├── docs/
│ ├── README.md
│ └── images/
└── report/
├── README.md
└── Case_Study_Report_Data_Handling_and_Decision_Making.pdf
Source & method of acquisition:
Secondary dataset published on Kaggle, created through web scraping by PromptCloud and DataStock.
Representativeness:
- Large dataset representing a wide range of eBay product listings (focused on electronics/accessories).
- Useful for modelling behavioural patterns around trust signals (seller ratings) and downstream engagement (reviews).
Initial dataset size:
- 19,230 rows × 30 columns
Key study variables:
- Independent variable (X): Seller Rating
- Dependent variable (Y): Seller Number of Reviews
In this study, review volume is treated as a proxy for purchase activity, since reviews occur after purchase on structured marketplaces.
To prevent high memory usage, the dataset was read in chunks (100 rows per chunk) and combined using concat() after iteration.
Why this matters:
Chunking supports stable processing on limited RAM machines and provides a scalable method for large datasets.
Key preparation actions:
- Removed unused/unrelated columns (retaining only study variables for modelling).
- Removed missing values from Seller Rating and Seller Num of Reviews (dropna).
- Ensured numeric type consistency before modelling.
Final modelling dataset size:
- 11,018 rows × 2 columns
From the cleaned dataset (N = 11,018):
Seller Rating
- Mean: 0.985
- Median: 0.987
- Std. Dev: 0.0102
- Min: 0.943
- Max: 1.00
Seller Num of Reviews
- Mean: 779
- Median: 354
- Std. Dev: 975
- Min: 0
- Max: 4,121
These distributions show that seller ratings are tightly clustered close to 1.0, while review counts are heavily right-skewed (many sellers with low reviews and a minority with very high engagement).
Regression was selected because:
- The study focuses on one predictor (seller rating) and one outcome (review volume).
- It supports inference (testing relationships) and forecasting (predicting likely review volume under different rating levels).
Scatter plots showed a clear non-linear trend: review volumes rise sharply as ratings approach 1.0.
Therefore, the modelling assumption was an exponential form:
[ y = a \cdot b^x ]
Where:
- (x) = seller rating
- (y) = seller number of reviews
- (a, b) = model constants
R² = 0.9913
This indicates the exponential model explains about 99.13% of the variability in review volume using seller rating alone — an exceptionally strong fit for behavioural marketplace data.
- At low rating levels (example: x = 0.971), predicted reviews remain very small (example point shown around y = 16).
- As seller rating approaches x = 1.0, review volume increases rapidly, reaching observed maxima such as y = 4,121 reviews.
This supports the interpretation that small improvements near high ratings can yield disproportionately large increases in customer engagement and purchasing trust signals.
Decision: Existing seller ratings influence customer purchasing behaviour.
Type: Strategic decision.
Recommended actions:
- Improve customer service responsiveness (fast replies, dispute resolution).
- Focus on product quality consistency (reduce returns and negative feedback).
- Provide seller training and best-practice guidance to sustain high rating performance.
- Monitor rating trends and intervene early for sellers drifting below the high-performing range.
Note:
Seller rating is not the only factor influencing purchases. Future modelling improvements should incorporate other explanatory variables such as price, product category, delivery time, and return rate.
- Python (data loading, filtering, manipulation, and plotting)
- Typical stack used in the report: pandas, NumPy, Matplotlib (and related libraries)
- Jamovi (descriptive statistics, box plots, outlier inspection)
- Desmos (equation modelling, exponential curve inspection, forecasting interpretation)
- Start with /report to read the full academic write-up.
- Use /docs/images to view the visuals referenced across the report:
- missing values
- outlier plots
- scatter plots
- exponential growth and forecast graphs
Samuel Boadi Agyekum
MSc Data Analytics & IT Security Management — Arden University
GitHub: https://github.com/agyekumboadi