30-Day Analyst Portfolio Challenge: Project Bank for Every Profession

"I analysed the Superstore dataset using Power BI. The business problem was finding which products and regions were actually profitable versus just high-revenue. I used DAX measures to calculate profit margins by category and found that Technology had a 17% margin while Furniture was dragging overall profit down at 4%. I recommended prioritising Technology in the West region."

"I used Python to build a churn prediction model on the IBM Telco dataset. The key insight was that contract type was the strongest predictor — month-to-month customers churned at 42% vs 11% for annual contracts. I built a logistic regression model and then visualised the high-risk segment in Power BI so the marketing team could use it without needing to code."

"I queried the Olist e-commerce database using SQL — joins, window functions, and CTEs to calculate rolling delivery times. The business problem was high late-delivery rates. I found the NE region was the main culprit — carrier switching mid-route was causing delays. I summarised this in an Excel pivot for the operations team."

"I exported Google Analytics data and built a funnel dashboard in Power BI. My key finding was that organic search users spent 4 minutes on site vs 90 seconds for paid social — and converted at 3.2% vs 0.9%. The marketing team was over-spending on Instagram ads that weren't converting."

"I exported 6 months of JIRA data and built a cycle time analysis in Excel. The team thought the developers were slow — but the data showed stories sat in the QA column for an average of 3.2 days. The fix wasn't hiring more developers. It was shifting QA to start testing earlier in the sprint."

"I ran a market basket analysis using Python's mlxtend library. The most interesting finding was that customers who bought 'JUMBO BAG RED RETROSPOT' also bought the blue version 73% of the time — but they weren't displayed together on the site. A simple cross-sell recommendation would capture that."

"The project was about designing for a non-technical stakeholder. I focused on layout first — the CEO needed to see green/red status in under 30 seconds. I used conditional formatting, sparklines, and card visuals. The data model behind it was simple but the UX design is what made it valuable."

"I built an ARIMA forecasting model in Python using public quarterly revenue data from Screener.in. The challenge was seasonality — the company had a strong Q4 spike every year. Once I decomposed the seasonal component, the model's MAPE dropped from 11% to 4.2%. The output was a 12-month forecast with confidence bands."

"The finance team was doing this manually every month — copying and pasting from four Excel files. I connected Power BI to a shared Excel model, built a variance table with conditional formatting for red/amber/green, and added a drill-down by cost category. The CFO could now answer 'why are we over budget in Marketing?' in 30 seconds."

"I downloaded 3 years of NSE stock data using the yfinance library. The investor thought they were diversified across IT, banking, and pharma — but the correlation matrix showed 7 of their 10 stocks moved almost identically. Their Sharpe ratio was 0.6, which is below what you'd expect for the risk they were taking."

"I built an attrition dashboard on the IBM HR dataset. The surprising finding was that it wasn't the employees with bad performance ratings leaving — it was mid-performers in the ₹3–5L salary band. They had enough skills to get hired elsewhere but weren't getting salary reviews. The intervention is a mid-band salary audit, not a performance improvement programme."

"The hiring team assumed candidates were dropping out after offers. But the funnel data showed the drop happened between phone screening and technical interview — a 12-day average wait. Candidates were taking other offers. The fix was simple: add two more technical interviewers and pre-schedule slots."

"RFM stands for Recency, Frequency, and Monetary — three signals that predict customer lifetime value. I scored each customer 1–5 on each dimension and created 5 segments. The key insight was that 23% of customers hadn't purchased in 90 days but had high historical spend — a classic at-risk-but-valuable segment. A targeted win-back campaign is much cheaper than acquiring new customers."

"The marketing team was ready to roll out the new page because it 'looked better' and had a higher conversion rate in the test. But I ran a chi-square test and found the p-value was 0.031 — statistically significant. More importantly, I also calculated the minimum detectable effect and confirmed the sample size was adequate. Without that check, a lot of A/B tests give false positives."

"I built a Power BI dashboard that connected to a shared Excel CRM export. The manager's key question was 'where should I focus coaching this week?' So I built a conditional heatmap — green for above 80% quota, amber for 60–80%, red for below 60%. At a glance, the South territory was consistently red and it turned out they had the lowest lead quality, not the weakest reps."

"The problem was that the team counted every deal in 'Proposal Sent' as 50% likely to close — but historical data showed only 22% of those actually closed. I built a weighted model using actual win rates per stage and recalculated the pipeline. The base case forecast was 30% lower than the VP's number — and it was much more accurate."

"I used Facebook's Prophet library to forecast demand for 500 SKUs across 6 stores. The insight was that the company was using a simple average-demand reorder model — which works fine for stable products but fails completely for seasonal ones. After applying Prophet's seasonality decomposition, the forecast accuracy improved from 68% to 91%."

"The procurement team was renewing carrier contracts based on relationships, not data. I built a weighted scorecard — 40% weight on cost, 40% on on-time delivery, 20% on damage rate. Carrier B was 22% cheaper and only 2 percentage points behind on delivery time. That data shifted the negotiation completely."

"I used the UCI Diabetes Readmission dataset. The key finding was that patients who had 3 or more prior emergency visits and were discharged on insulin had a 34% readmission rate — vs 12% for the rest. A simple flag in the discharge system to trigger a 48-hour follow-up call for this group would significantly reduce readmissions."

"I used the Our World in Data vaccination dataset and built a choropleth map in Power BI to show coverage by state. The pattern was clear — low coverage correlated strongly with rural population density, not vaccine hesitancy as was assumed. The intervention should be access, not messaging."

"I used SQL to query event logs and built a funnel from install to first transaction. The funnel looked fine until step 4 — KYC document upload. That one step had a 62% drop-off rate. Users were abandoning because the form required both PAN and Aadhaar uploads immediately. Moving KYC to after first login, not before, would keep users engaged while meeting regulatory requirements."

"I used SQL to calculate feature adoption rates broken down by churned vs retained users. The 'north star feature' pattern was clear — users who activated the report export feature within the first 7 days had 4× higher 90-day retention. The product team was spending roadmap budget on new features when they should have been making report export easier to find in onboarding."

"Price elasticity measures how sensitive demand is to price changes. I used log-log regression to estimate elasticity coefficients for each product category. Electronics had an elasticity of -2.3 — meaning a 10% price increase drops demand by 23%. Books were -0.4 — nearly inelastic. That means you can raise book prices by 10% and only lose 4% of volume, which is net positive margin."

"The returns problem wasn't uniform — it was concentrated. Three sellers in the Fashion category accounted for 41% of all returns, and 'size mismatch' was the top return reason for all three. The fix wasn't better reverse logistics — it was adding a mandatory size chart requirement for these sellers in the Fashion category."

"I built the scorecard using the standard WoE + logistic regression approach that most banks use for retail credit. The Gini of 0.61 is considered acceptable for a retail loan portfolio. The key finding was that external credit bureau score was the single strongest predictor — but debt-to-income ratio caught a segment of borrowers with good scores who were still high-risk."

"Fraud detection is an imbalanced classification problem — fraud is only 3% of claims. I used SMOTE to oversample the minority class and then trained an XGBoost model. At a 5% flagging threshold, the model had 89% precision — meaning 89 out of every 100 flagged claims had genuine fraud indicators. The network analysis showing repeat claimants using the same workshops was the most actionable finding."

"The strongest signal for dropout wasn't exam performance — it was forum participation in Week 1. Students who posted even one comment in the first week had a 67% completion rate vs 28% for those who never posted. The intervention is simple: a personal message from a tutor to any student with zero forum activity in the first 5 days."

"I used Python's VADER sentiment library — pre-built for social media text — to classify 50,000 posts. Rather than just positive vs negative counts, I used topic modelling (LDA) to find what the negative posts were actually about. Packaging was the dominant theme, not the product formula. The brand team thought the formula change was controversial — the data showed it was the new box design that consumers hated."

"I built a crop recommendation model using soil and climate features. The accuracy was 98% using Random Forest — but the more valuable output was the feature importance analysis showing that soil pH was the most important factor, not rainfall as commonly assumed. This changes how extension officers should prioritise soil testing in their field visits."

"The data showed a clear pattern — OTA bookings with more than 60 days lead time cancelled at 52%. Direct bookings under 14 days cancelled at only 8%. The revenue strategy should be to push direct booking incentives (free breakfast, room upgrade) for the 60+ day advance window and accept OTA bookings with a non-refundable deposit for that lead time."