What Makes a Call Fraudulent? Six Key Warning Signs

By shrijeetverma13 · June 25, 2026

This analysis identifies the strongest behavioral and account indicators that separate fraud from legitimate calls. We found that brand-new accounts, high…

Fraud probability sharply increases at a country code risk score of approximately 52. At this threshold, the fraud rate jumps by about 17.2 percentage points compared to the previous bin. A chart has been generated showing the fraud rate curve across all risk score ranges, with a red dashed line marking this critical threshold.

Yes, newer accounts are strongly associated with higher fraud rates. The analysis reveals a clear and dramatic relationship: accounts aged 0-30 days have a 100% fraud rate, while the oldest accounts (1,826-3,650 days) show a 0% fraud rate. The correlation between account age and fraud is -0.463, confirming a meaningful negative relationship — as account age increases, fraud likelihood decreases significantly.

The analysis examined how fraud rates change across different spam report count buckets. A bar chart was generated showing the fraud rate percentage for each bucket (0, 1-2, 3-5, 6-10, 11-20, 21-50, and 50+ spam reports), along with supporting data tables.

Yes, both calls per day and unique receivers 24h are strong indicators of elevated fraud risk. The analysis shows a very strong positive correlation for both metrics — 0.752 for calls per day and 0.759 for unique receivers 24h — meaning users with higher values in either metric are significantly more likely to be flagged as fraudulent. Two bar charts were generated showing how fraud rates rise sharply across quartiles for both metrics.

The analysis reveals a strong relationship between fraud rates and both reputation score and previous fraud associations. Entities with the lowest reputation scores (0-20) have a 100% fraud rate, while those with the highest scores (80-100) have almost no fraud at just 0.03%. Similarly, entities with 21 or more previous fraud associations show a 100% fraud rate, compared to only 5.69% for those with none. Three charts and supporting data tables were generated to visualize these patterns, including bar charts for each dimension and a heatmap showing the combined effect of both variables.

The analysis identified the behavioral metrics that most strongly separate fraudulent from legitimate calls. Two bar charts were generated showing effect sizes (Cohen's d) and correlations with the fraud label across 13 metrics. The top differentiators are receiver block rate, sequential dialing score, and reputation score — all with very large effect sizes and correlations above 0.86.

The analysis segmented callers into 4 quartile groups based on call duration (Q1=shortest, Q4=longest) and produced a grouped bar chart comparing average call duration, block rate, spam reports, and fraud rate across each quartile. The visualization and data tables are available showing how these key metrics vary across the four duration segments, helping identify whether call length is a meaningful fraud risk indicator.

The IQR outlier detection analysis was performed on both receiver block rate and graph degree columns, and box plots were generated for each column split by fraud label (Fraud vs. Non-Fraud). The visualizations clearly highlight the distribution of values and where outliers fall for each group, making it easy to compare outlier patterns between fraudulent and non-fraudulent transactions.

A full Pearson correlation matrix was computed across all 14 numeric features and visualized as an annotated heatmap. A second bar chart highlights the top features most strongly correlated with fraud label. Both visualizations are available for exploration.

The analysis reveals striking differences between fraudulent and legitimate callers across both behavioral metrics. Two overlapping histogram charts with KDE curves were generated showing the distribution shapes for each group, along with detailed statistics tables.

For night call ratio , fraud cases average 57.76 (median 58.33, std 17.57) compared to just 19.88 (median 17.63, std 12.04) for non-fraud — nearly 3x higher. For sequential dialing score , fraud averages 71.59 (median 73.38, std 15.82) versus only 16.60 (median 14.77, std 10.25) for non-fraud — over 4x higher. Both differences are statistically significant with a Mann-Whitney p-value of essentially 0, confirming these are not random variations.