Tool-Induced Analytical Variance in Credit Risk Exploration: A Comparative Study Using Python, SQL, and Power BI

Financial data analysts commonly use Python, SQL, and Power BI to analyse financial datasets, yet little formal examination evaluates whether these tools produce reliably consistent results on the same data. Inconsistencies between tools could affect analytical quality and lead to contradictory business decisions a critical concern in credit risk contexts where decisions directly affect borrowers and creditors. While larger financial institutions rely on standardised data pipelines, smaller lending organisations, fintech startups, and private credit providers frequently analyse data in tool-specific environments, making tool-induced variance a practically significant but underexplored problem. Machine learning models for credit risk prediction have received considerable research attention, yet the upstream analytical pipeline how tool choice affects data handling and insight generation remains understudied.

This study applies exploratory data analysis across Python, SQL, and Power BI on the Give Me Some Credit dataset (150,000 borrowers, Kaggle). A key finding emerged during data preparation: missing values encoded as "NA" were handled inconsistently. Python's pandas correctly identified 29,731 missing income values as NaN. SQLite's CAST function silently converted "NA" to zero an invisible data integrity risk. Power BI imported the same column as Text type, requiring manual intervention before numerical analysis was possible.

Four findings were confirmed across all three tools: a 93%/7% class imbalance between non-defaulters and defaulters; defaults peaking in the 40-60 age group; debt ratio emerging as a surprisingly weak predictor; and late payment history as the strongest predictor, with defaulters averaging 15 times more 90-day late payments. Relying solely on debt ratio would have produced misleading conclusions, demonstrating the risk of single-tool analysis.

These findings support standardised upstream data preparation pipelines particularly for fintech startups and small credit institutions without dedicated data engineering teams, where tool choice can silently affect credit decision quality.