Projects

Ridge and Lasso regression to predict house prices for a company buying undervalued homes in Australia, on a 1,460-row / 81-feature dataset. Lasso at the optimal alpha (0.0001) was chosen as the final model — it zeroed out most coefficients, cutting the effective feature count to 154 while matching Ridge's train/test R², giving a simpler and more generalizable model plus a ranked list of the variables that actually drive price.

PCA followed by K-Means and hierarchical clustering on socio-economic and health indicators for 167 countries, to group nations by overall development for an NGO deciding where aid money matters most. K=2 gave the best silhouette score (0.50), but a finer split was used for the final segmentation so the low-development cluster wasn't so broad it stopped being actionable.

Exploratory analysis of a lending company's applicant data (307,511 rows, 122 columns) to understand which applicant attributes correlate with payment difficulty. Most of the work is data cleaning — dropping columns over 50% missing, deciding what can and can't be safely imputed for categorical vs. numeric fields — before getting to the actual default-risk patterns.

Analysis of Indian state-wise GDP growth data, merged with education drop-out statistics to test whether states with lower per-capita GDP see more students drop off before finishing school. The data confirmed it — per-capita GDP is inversely correlated with Primary and (most strongly) Upper Primary drop-out rates, tying low education completion to states' share of GDP from manufacturing and services versus agriculture.

Logistic regression to score ~9,000 inbound leads by conversion likelihood for an online course provider (X Education), so sales can focus on the leads actually worth calling. The final model reached about 92% train accuracy and 90.7% test accuracy at the chosen probability cutoff, with sensitivity/specificity tradeoffs plotted across cutoffs to pick the one matching the business's target conversion rate.

Multiple linear regression to help a US car manufacturer understand what drives price in the Chinese market, distilled down from the full feature set to a 6-variable final model (via RFE and manual VIF/p-value pruning). That model holds up well out of sample — R² of about 89.6% on train and 89.9% on test, adjusted R² around 89.1%.

Predicting churn among a telecom's high-value customers, comparing logistic regression (with SMOTE for class imbalance) against a grid-search-tuned random forest. Logistic + SMOTE gave the best balanced result (~82% test accuracy, ~0.84 AUC) and stayed the primary model since it directly surfaces which features matter — the random forest edged out on raw accuracy (~94%) but only caught 46% of actual churners, a worse trade for a business that wants to act on the prediction. Roaming usage, total minutes of usage, ARPU, and max recharge amount were the strongest churn signals.

A separate, individual pass at telecom churn prediction using multiple logistic regression, with feature selection via RFE and multicollinearity checked through VIF. After tuning the classification cutoff against the sensitivity/specificity tradeoff, the final model landed at about 77% test accuracy.

A classic inferential-statistics project: a paired one-tailed t-test on reaction times from the Stroop task (naming ink colors when the word and color conflict). Mean reaction time was 14.05s for congruent words vs. 22.02s for incongruent ones — a large, statistically significant slowdown, confirming the interference effect the task is named for.

Exploratory analysis of the Kaggle Titanic passenger dataset to find what predicted survival. Class, sex, and port of embarkation stood out as real factors — being female and in 1st class gave the best odds, a male in 3rd class the worst — while age alone wasn't a strong predictor, and having family aboard was linked to modestly better survival odds.