SHANEEZA HASNANI CFE

FRAUD ANALYTICS · DATA SCIENCE

I follow the data
until the story
makes sense.

I’m Shaneeza Hasnani, a Certified Fraud Examiner who uses Python, R, SQL, and machine learning to investigate financial crime and reduce false positives.

Read the project stories ↓Open my résumé
Hand-drawn illustration of Shaneeza seated on a marble staircase in her graduation stole
I like the point where
messy data becomes a useful decision.
CERTIFIED
FRAUD EXAMINER
ABOUT / 01

Data scientist.
Fraud examiner.

I study how suspicious behavior shows up in data, then build practical ways to find it.

PROJECT CASE STUDIES / 02

From question
to evidence.

Each card tells the short version. The repositories hold the code, models, tests, and methodology.

01

FRAUD DETECTION · SNOWFLAKE + SQL

Turning 6.3 million transactions into a review queue

The problem

PaySim contains millions of mobile-money transactions, but only 0.13% are fraud. A useful system had to find the risky behavior without treating every unusual payment as equally suspicious.

What I tried

I built raw, staging, and mart layers in Snowflake. Then I tested three behavioral signals: an origin account dropping to zero, a destination balance staying unchanged after receiving funds, and transaction types associated with every fraud case in the dataset.

What happened

The scoring rules sorted every transaction into four risk tiers. The HIGH tier contained 5,145 transactions, and 76% were labeled fraud. That gave an investigator a much smaller place to start.

What I’d test next

PaySim is synthetic. With real bank data, I would test stability over time, monitor feature drift, and measure how many legitimate customers the rules inconvenience.

Technical details on GitHub ↗
Snowflake fraud detection pipeline
6.3M rows · 3 behavioral features · 4 risk tiers
02

MARKET EVENT STUDY · PYTHON

What happens after corporate fraud becomes public?

The question

I wanted to separate the market reaction to a fraud disclosure from normal market movement.

What I did

I curated 25 confirmed disclosures and used the Fama-French three-factor model to estimate expected returns before measuring abnormal returns around each event.

What I found

Cumulative average abnormal returns reached roughly −20% by day +10. The continued decline suggested that investors were still reassessing the company after disclosure day.

The limit

The sample is small and intentionally focused on high-profile cases. Other company news may overlap with some disclosure windows.

Methods and tests on GitHub ↗
CAAR AFTER DISCLOSURE
DISCLOSURE
Replace this illustration with the final repository chart.
03

DATA ENGINEERING · PYTHON + SQLITE

Building the table a fraud model needs

The problem

Fraud demonstrations often begin with a clean file. Production work begins with duplicates, missing values, invalid amounts, and inconsistent records.

What I built

I generated 10,000 synthetic transactions, injected realistic data-quality problems, cleaned them, and created five fraud signals plus a composite risk score.

The result

The pipeline turned 10,100 raw rows into 9,753 validated records, a SQLite database, and a separate alert file. Twenty-seven tests check the workflow.

Why it is modular

The extract, transform, and load stages are separate, so a real API or warehouse can replace the synthetic source without rewriting the feature logic.

Pipeline and tests on GitHub ↗
Fraud transaction ETL pipeline
10,100 raw rows → 9,753 validated rows
04

CLASSIFICATION · PYTHON

Starting with an interpretable baseline

The question

Which passenger attributes best explain Titanic survival, and how far can a simple logistic regression go?

The approach

I cleaned missing values, encoded categorical variables, split the data 80/20, and chose logistic regression so the prediction remained explainable.

The result

The model reached about 78% test accuracy. Gender, passenger class, and fare were influential, but accuracy alone leaves important classification questions unanswered.

Model details on GitHub ↗
05

TRANSIT ANALYSIS · R

When two Metro stations need different plans

The problem

A transit schedule built around an average day can miss how demand changes by station, weekday, and holiday.

What I compared

I analyzed Monday and Friday ridership at Foggy Bottom-GWU and Tenleytown-AU from August 2021 through July 2024.

What stood out

Fridays were consistently busier, holiday Mondays dropped sharply, and Foggy Bottom had both higher volume and more variation than Tenleytown.

Analysis and visualizations ↗
06

COMPENSATION ANALYSIS · R

Testing whether career level tracks salary in NYC

The problem

Job titles are inconsistent, and posted pay can be hourly, daily, or annual. Comparing compensation required a common scale.

What I changed

I standardized wages, calculated midpoint salaries, grouped more than 5,300 listings into career levels, and tested the relationship with salary range.

The result

The association was statistically significant with a moderate effect size (Cramer’s V = 0.28). Leadership roles concentrated in the highest salary group.

Statistical workflow on GitHub ↗
07

LABOR-MARKET ANALYTICS · R

Skills appear in bundles, not isolation

The question

Which skills appear together in LinkedIn postings, and how do those combinations relate to industry, experience, and salary?

What I tried

I joined job, skill, salary, and industry tables, then combined association rules, clustering, decision trees, and regression.

What emerged

Distinct skill-based job profiles had meaningful salary differences. Experience increased salary, but the size of that increase depended on industry.

Full analytical workflow ↗
08

AUDIO ANALYTICS · R + QUARTO

What makes a song sound energetic?

The question

I tested how loudness, tempo, and genre relate to the energy score assigned to Spotify tracks from the 2000s.

The approach

I cleaned the audio features, grouped genres, built a multiple regression model, and checked its assumptions and statistical power.

The finding

Loudness was the strongest predictor, while tempo had a smaller positive relationship. The model explained about 54% of the variation in energy.

Report and model diagnostics ↗

EXPERIENCE / 03

Work that had to support a decision

These details come from the résumé attached for this portfolio.

Research Assistant

Schmidt Special Competitive Studies Project

  • Combined Airtable, Google Sheets, and third-party APIs into a company-research dashboard that cut meeting preparation time by 50%.
  • Automated sponsor-news collection, saving three hours of manual research each week.
  • Tracked actions across more than 30 sponsor companies and prepared weekly leadership summaries.

Fraud Data Analyst

EduGuide Overseas

  • Built anomaly-detection models in Python and R and classified more than 200,000 transactions by fraud risk each month.
  • Analyzed more than 500 payment datasets monthly and reduced suspicious student applications through automated scoring.
  • Built Power BI reporting that cut review time by 15%.

Financial Crime & Investigative Services Intern

Guidehouse

  • Analyzed more than 20,000 ACH and wire transfers with SQL and Python.
  • Engineered transaction features and identified more than 50 high-risk patterns for compliance investigation.

Fraud Audit Intern

New York State Office of the Attorney General

  • Used SQL and statistical validation to identify more than $500,000 in recovery anomalies.

EDUCATION + CREDENTIALS / 04

DEC 2026

M.S. Business Analytics & AI

Data Science and Statistics concentration
American University

MAY 2025

B.S. Fraud Examination & Financial Forensics

Computer Science minor
John Jay College

CREDENTIALS

Certified Fraud Examiner

UiPath Agentic Automation Developer Associate

CONTACT / 05

If the work involves fraud, risk, or difficult data, I’d like to hear about it.

hasnani.shaneeza@gmail.com
LinkedIn ↗GitHub ↗Résumé ↓