What should a new Amazon store sell first? We used prices and ratings from a 2023 Amazon India catalog snapshot to compare store ideas and build a list of products a seller could research before buying stock.
UT Dallas / Business Analytics with R / Spring 2025
Our recommendationStorefront to research
Home organization
01Home storageCloset bags and drawer dividers
02Kitchen storageAirtight pantry containers
01 The problem
Why we did it
A new seller has to pick a niche and commit money to inventory before knowing what will sell. With so many products already on Amazon, we wanted a way to narrow the field and decide what was worth looking into first.
02 The approach
How we did it
01
Clean the catalog
We cleaned prices and ratings in R. Of the 551,585 listing rows, 362,797 had enough data to model.
02
Score and test products
We scored each listing, then tested whether a model could recognize the listings we had scored highest.
03
Compare store ideas
We ran the same screen across six categories, two per store idea, and counted how many listings passed.
The model experiment
Could a model spot the products we scored highest?
We called a listing high-scoring when it met the top 10% cutoff from the training data. Then we tested how well a decision tree found those listings on products it hadn’t seen.
How we scored each listing
Price×Rating count×Stars
We used discounted price where available and list price elsewhere. Rating count stood in for orders.
Inside the fitted tree
Follow one branch
The tree split listings into groups using simple rules. This branch starts with two rules from the all-input model.
Rating count833 or more
Listing price₹507.50 or more
Does it have at least 1,223 ratings?
Yes
Predict high score
99.0%
12,480 of 12,606 training rows scored high.
No · 833–1,222 ratings
Check the price again
Below ₹768.50
Predict lower score
₹768.50 or more
Predict high score
Other listings follow other branches. The percentage describes the training rows in this group.
Model check · Decision trees
How much did rating count matter?
We tested both trees on the same 72,531 listing rows. The products in this set were held out of training.
Recall is the share of the 7,377 high-score test rows the model found.
Price, stars, and rating count97.10%
7,163 of 7,377
Price and stars24.48%
1,806 of 7,377
0%25%50%75%100%
See accuracy, precision, and the baseline
Other metrics on the same test set
Metric
With rating count
Without rating count
Accuracy
99.39%
91.05%
Precision
96.95% 7,163 of 7,388
66.23% 1,806 of 2,727
ROC AUC
0.9971
0.6801
Accuracy is out of all test rows. Precision is out of the rows each model predicted as high-scoring. ROC AUC describes how well the model ranked high-score rows ahead of lower-score rows.
Always predicting the lower-score group gave 89.83% accuracy and caught zero high-score rows.
Rating count was part of the score itself. Applying the formula directly found all 7,377 high-score test rows by definition. To test sales or profit, we’d need transaction and cost data.
The model could learn our score. A seller still needs a way to compare the options in each category, so we put the same filters on all three store ideas.
See the catalog breakdown and four-model comparison
Source rows
551,585
Product IDs you can browse
471,156
Subcategories with listings
112
Data preparation
Which listings had enough data?
We brought together 139 files and cleaned up the prices and ratings.
Kept for analysis
362,797
Missing both rating fields
175,794
Invalid rating or rating count
6,233
Missing or unreadable list price
6,761
We could use 65.8% of rows to train or test a model. The explorer still shows listings with missing fields.
Catalog coverage
How many listings have ratings?
How often a listing has a star rating, by subcategory.
Gold & Diamond Jewellery8.0%
Air Conditioners58.8%
Headphones71.8%
All Electronics98.8%
Among rated listings, Gold & Diamond Jewellery and All Electronics both had a median of 4.1 stars. The median rating count was 2 for jewellery and 1,174 for electronics.
Four-model comparison
We tried three approaches. Logistic regression estimated how likely a listing was to have a high score. Classification trees (CART) split listings with yes-or-no rules. K-nearest neighbors looked at similar listings and used their labels to make a prediction. We tuned one tree through cross-validation and ran K-nearest neighbors on a smaller sample.
Share of high-scoring listings each model found, as reported in the project presentation
Logistic regression84.91%
Untuned tree84.04%
Tuned tree90.44%
K-nearest neighbors82.12%
Results reported in the project presentation
Model
Accuracy
Recall
Specificity
Precision
ROC AUC
Logistic regression
95.48%
84.91%
96.37%
65.89%
0.906
Untuned CART
91.25%
84.04%
91.85%
34.25%
0.879
Tuned CART
98.38%
90.44%
99.28%
93.49%
0.950
K-nearest neighbors
95.57%
82.12%
96.84%
71.23%
0.889
The four results above come from the project presentation. The tree experiment shown earlier used a separate product-level test set. In this table, the reported precision for logistic regression and untuned CART does not line up with the other metrics if they came from one set of predictions.
Example picks
Five of the tuned tree’s ten displayed picks were air conditioners, including LG and Voltas listings. Across the three model lists, only the Samsung 980 1TB SSD and the Prestige Electric Kettle PKOSS appeared twice.
Logistic regression
Amazon Basics 6 Foot Micro…
Samsung 980 1TB Up to…
Tuned CART
LG 1.5 Ton 5 Star…
Voltas 1.4 Ton 3 Star…
K-nearest neighbors
Hammonds Flycatcher Genuine Leather Wallets…
GreenLoop - All In One…
03 What we found
Home organization was the strongest starting point
About 30% of Home Storage listings and 43% of Kitchen Storage & Containers listings passed our screen. Both categories had more matches than any category in the travel bags or home fitness ideas, even when we raised the minimum to 100 ratings.
Catalog screen
Shortlist matches by storefront
Observed price above ₹0 and up to ₹3,500, with at least 4 stars.
Ratings
0300600900
Home organization
25+ ratings
801
100+ ratings
544
Travel bags
25+ ratings
174
100+ ratings
121
Home fitness
25+ ratings
203
100+ ratings
123
Each total adds the matches from two categories. A listing in both categories can count twice.
Why more matches help. Home organization gives us more examples to compare on price and features across both product lines. A deeper pool can also mean more competition, so we’d check recent sales, leading sellers, and margins before buying inventory.
Three products to research
We chose a closet, drawer, and pantry product to give the storefront a focus. For each type we took the listing with the most ratings among screened matches to a product name search.
Closet
GONGSHI vacuum storage bags
₹1,599 · 4.4 stars · 16,743 ratings
Drawer
AB SALES drawer dividers
₹499 · 4.1 stars · 1,115 ratings
Pantry
Solimo airtight pantry jars
₹489 · 4.2 stars · 2,809 ratings
Prices and ratings are from the catalog snapshot. We searched listing titles for “storage bag”, “drawer organi” (to catch both spellings of organizer), and “airtight jar”.
Why these three ideas? We picked three shopper needs: organizing a home, packing for a trip, and working out at home. Each idea has two related categories so we can check for options across more than one product line.
Why these filters? We used the same ₹3,500 price ceiling and 4-star minimum across all six categories. Requiring 25 ratings gave each average more than a handful of ratings to draw on. Raising the cutoff to 100 let us see whether the result held.
Matching listings by category
Category
25+ ratings
100+ ratings
Home organization
Home Storage
364 of 1,201
241
Kitchen Storage & Containers
437 of 1,011
303
Travel bags
Backpacks
144 of 2,062
103
Travel Duffles
30 of 1,152
18
Home fitness
Fitness Accessories
62 of 1,296
35
Strength Training
141 of 1,104
88
Each category counts distinct listings. Listings with missing prices or ratings stay in the total, so the results also reflect how complete the catalog is for each category.
We multiply price by rating count and stars to get the score. When the discounted price is missing, we use the list price. Rank is based on listings in the same subcategory with all three values. Tied scores share a rank.
We built a repeatable way to compare storefront ideas in R and turn the results into a shortlist. Of the three ideas, we’d research home organization first, starting with storage bags, drawer dividers, and pantry jars.
Before placing an order, a seller would check recent sales, supplier quotes, marketplace fees, shipping, and returns to estimate the margin.