Building an Interactive Excel Dashboard.

The Setup The dataset: 112 products scraped from Jumia Kenya, with fields for current price, old price, discount percentage, review count, and customer rating. The goal was to clean it, analyze it, and turn it into an interactive Excel dashboard - the kind with slicers and pivot charts, not just static tables. Simple enough on paper. The data had other plans. Cleaning: The Unglamorous Part That Actually Matters Before any analysis, I had to fix: - Prices stored as text - Excel treats "KSh 1,499" as a string, not a number, so no math works until you strip the formatting - Ratings buried in descriptive text - extracting a clean numeric rating from messy scraped fields - Invalid negative review counts - clearly a scraping artifact, not real data - 55 of 112 products (49%) with no review or rating data at all That last one was the decision point that shaped everything after it. I could have dropped those 55 rows. Instead, I labeled them explicitly - "No reviews" and "Not Rated" - rather than treating them as noise. That turned out to matter a lot, because half the catalog having zero customer feedback is itself one of the most important findings in the whole project, not something to clean away. Feature Engineering: Building Categories for Pivot Analysis To make pivot tables and slicers actually useful, I added four calculated columns: - Review_Category - Low / Medium / High / No reviews - Rating_category - Poor / Average / Excellent / Not Rated - Discount_category - Low / Medium / High Discount - price_category - Low / Medium / High / Premium (quartile-based bands) This is the part that felt most like actual data engineering rather than spreadsheet work - deciding on thresholds, handling edge cases (blank cells silently evaluating as 0 in Excel formulas, which almost mislabeled every unrated product as "Poor" quality), and making sure every downstream pivot table had clean, consistent categories to group by. Building the Dashboard I built 8 pivot tables from a single shared table (tblCleaned ), which turned out to be the key technical decision: every pivot table needs to come from the same source table for slicers to control all of them simultaneously. Miss this, and you end up with a slicer that only filters one chart instead of your whole dashboard - a mistake I made and had to fix along the way. The dashboard includes: - 5 KPI cards (total products, average price, average discount, average rating, total reviews) - 3 top-10 bar charts (by rating, by reviews, by discount) - 3 trend charts (discount vs. reviews, rating vs. reviews, price vs. rating) - 2 pie charts (rating breakdown, discount breakdown) - 3 slicers (rating category, discount category, price category) connected across all 8 pivots What the Data Actually Said Here's where it got interesting. I went in assuming discounts drive engagement - that's the whole premise of a sale, right? The data disagreed. Discount % vs. Review count: correlation ≈ -0.14 Higher discounts do not lead to more reviews. If anything, there's a very weak negative relationship. Rating vs. Review count: correlation ≈ +0.06 Essentially zero relationship. Highly-rated products don't get meaningfully more reviews than mediocre ones. Price vs. Rating: correlation ≈ +0.15 Barely positive. Expensive products are not reliably rated higher than cheap ones - though mid-to-upper-priced items (KSh 1,190-1,820) did show the best average rating (~4.2) of any price tier. The bigger story: - Only 16 of 112 products (14%) hit "strong engagement" (14+ reviews) - 90% of heavily discounted products (40%+ off) show low or no engagement - 49% of the entire catalog has zero reviews, regardless of discount or price The product that stuck with me most: a 120W Cordless Vacuum Cleaner with 69 reviews - the most of any product in the dataset - sitting at just 2.8★. High visibility, low satisfaction. That's a much bigger red flag than a discount that didn't move the needle. Takeaways for Sellers (and for Me) The clearest business insight: discounting isn't a substitute for product quality or visibility. A huge chunk of this catalog is heavily discounted with little to show for it, while a small number of genuinely well-reviewed products (crochet tool sets, cake decorating kits, a digital caliper) prove that solid ratings and real engagement can coexist without needing steep price cuts. For me personally, this project reinforced something I keep re-learning in this cohort: the "boring" cleaning and validation work is where the real insight-generation happens. The blank-cell formula bug that almost mislabeled unrated products as "Poor," the pivot cache mismatch that broke my slicers, the decision to keep unrated products visible instead of dropping them - none of that shows up in the final dashboard screenshot, but all of it is what made the dashboard actually trustworthy. Repo & Full Breakdown Full project, including the cleaned dataset, the Excel workbook (raw data → cleaned data → pivot tables → dashboard), and a detailed README with all findings: Top comments (0)