Presentation Script
Slide 1 – Title Slide
My project titled Predictive Analytics for Digital Commerce Performance and Sales Risk Using Machine
Learning. This project was developed as a business analytics solution in a TikTok Shop-style digital
commerce environment. It focuses on using machine learning to identify high-performing products and
also detect products that may be at risk of sales decline. The study used the TikTok Shop Seller
Performance Dataset and transformed it into an enriched dataset for predictive modelling.
Slide 2 – Business Problem & Project Objectives
The main business problem behind this project is that many online sellers already have access to
product, traffic, campaign, and sales data, but they still rely mostly on descriptive dashboards and past
reports. This means that important decisions such as advertising budget allocation, stock planning, and
campaign timing are often made reactively rather than predictively. The objectives of this project were
to predict high-performing products, identify products at risk of sales drop, analyse important business
drivers, explore time-based and campaign-based patterns, and provide practical recommendations for
sellers.
Slide 3 – Dataset & Enrichment
This project used the TikTok Shop Seller Performance Dataset obtained from Kaggle. The original dataset
contained 2,000 observations and 21 columns. After data cleaning, preprocessing, and feature
engineering, it was transformed into an enriched dataset containing 2,000 observations and 46
variables. The final dataset included original commerce fields as well as engineered variables such as
funnel ratios, engagement indicators, stock, ratings, campaign timing, and two target variables, which
are is_high_sales and is_sales_drop.
Slide 4 – Analytical Framework
The analytical framework followed a full business analytics workflow. It started with data collection and
understanding, then data cleaning, feature engineering, exploratory analysis, preprocessing, model
training, model comparison, evaluation, and finally business interpretation. Two supervised learning
problems were addressed. The first was high-sales prediction, which focuses on identifying strong-
performing products. The second was sales-drop risk prediction, which focuses on identifying products
that may require early intervention.
Slide 5 – Exploratory Analysis: Product & Category Patterns
In the exploratory analysis stage, I examined product categories, pricing patterns, and revenue
performance. The purpose was to understand which categories contribute most to total revenue and
how product-level performance differs across the dataset. The analysis showed that Electronics was the
top category by total revenue. This stage also helped identify the overall distribution of pricing and
conversion behaviour before moving into predictive modelling.
Slide 6 – Funnel, Advertising & Stock Risk Signals
This slide focuses on funnel and business-performance variables such as views, clicks, add-to-cart
actions, orders, ad spend, and stock level. These variables are important because they represent the
journey from product exposure to purchase. The analysis also showed that low-stock products had the
highest sales-drop risk. This is business-relevant because it suggests that inventory condition is strongly
connected to performance sustainability.
Slide 7 – Time-Based and Campaign Insights
This slide presents time-based and campaign-related findings. The analysis identified Wednesday as the
strongest day by average conversion and 11:00 as the strongest hour by average conversion. Campaign
timing was also analysed by comparing periods such as morning peak, evening peak, weekend, and off-
peak. It is important to note that the day-of-week and hour-of-day variables were simulated because the
original dataset did not contain actual transaction timestamps. Therefore, these findings are used as an
academic demonstration of time-based analysis.
Slide 8 – Model Design: Six Algorithms Compared
To strengthen the modelling process, six machine learning algorithms were compared. These were
Logistic Regression, Random Forest, Decision Tree, Gradient Boosting, K-Nearest Neighbors, and Support
Vector Machine. Since the project had two target variables, this resulted in a total of twelve model runs.
This comparison allowed the best model to be selected based on performance rather than assumption.
Slide 9 – Model Performance Comparison
This slide compares model performance across the six algorithms. The models were evaluated using
accuracy, precision, recall, F1-score, ROC-AUC, and average precision. The results clearly show that the
best model differs depending on the business task. This is important because no single algorithm
performs best for every problem. The comparison demonstrates why multiple algorithms should always
be tested in business analytics projects.
Slide 10 – High-Sales Prediction: Champion Model
For the high-sales prediction task, the champion model was Gradient Boosting with one-hot encoding
and median imputation. It achieved an accuracy of 0.966, precision of 0.929, recall of 0.960, and an F1-
score of 0.944. This means the model was very strong at identifying products that are likely to perform
well. From a business perspective, this model can help sellers decide which products deserve more
advertising support, stronger campaign placement, stock protection, and content prioritisation.
Slide 11 – Sales-Drop Risk Prediction: Champion Model
For the sales-drop prediction task, the champion model was Support Vector Machine with
StandardScaler. It achieved an accuracy of 0.776, precision of 0.233, recall of 0.698, and an F1-score of
0.349. Although the performance is lower than the high-sales task, this is expected because sales-drop
prediction is a more difficult and imbalanced problem. Even so, the model is still useful because it can
serve as an early warning system for sellers, helping them review products before a bigger decline
becomes visible.
Slide 12 – Model Confidence and Diagnostic Curves
This slide includes ROC curves, precision-recall curves, probability distributions, and actual-versus-
predicted class counts. These diagnostic visuals help explain the reliability of the models beyond simple
accuracy. The high-sales model shows much clearer separation between the two classes, which supports
its stronger performance. The sales-drop model shows a more mixed pattern, which reflects the greater
complexity of risk prediction. This is why F1-score and recall were especially important in model
selection.
Slide 13 – Business Recommendations for Sellers
Based on the results, several recommendations were developed. Sellers should prioritise products with
strong add-to-cart activity, clicks, review count, and engagement. Advertising budget should be
allocated more carefully toward products showing stronger funnel quality. Low-stock products should be
reviewed more closely because they are more associated with sales-drop risk. Sellers should also
consider campaign timing and use model-generated risk flags as part of routine performance
monitoring.
Slide 14 – Limitations and Future Work
This project has some limitations. The original dataset was obtained from Kaggle, and the enriched
dataset contains engineered variables to support the project objectives. The target variables were also
created using business logic rather than a live platform label. In addition, the analysis is at product level
and does not include customer-level demographics, competitor prices, delivery speed, or platform
algorithm changes. In future work, the project could be extended using real platform data, more
advanced dashboards, time-series forecasting, profit-based optimisation, and additional algorithms such
as XGBoost or LightGBM.
Slide 15 – Final Takeaways
To conclude, this project shows that predictive analytics can improve decision-making in digital
commerce by transforming product-performance data into useful business insight. The project
successfully combined exploratory analysis, feature engineering, six-model comparison, evaluation, and
business interpretation. The final output is not only academically valuable but also practically useful for
sellers who want to move from reactive decisions to data-driven strategy.