Dataset Background

This dataset is a hybrid — it combines a real-world tax audit dataset with synthetically generated tax fraud cases.

The base audit records reflect genuine filing behaviour: income levels, deduction patterns, demographic attributes, and filing habits drawn from real audit data. However, the original audit dataset did not contain a sufficient number of confirmed fraud cases to train a supervised fraud detection model. To address this, synthetic fraud examples were generated and added to the dataset, simulating the kinds of anomalous patterns — inflated deductions, unusual credit claims, atypical filing behaviour — that are associated with fraudulent returns in practice.

The result is a dataset that preserves realistic distributional properties from real audits while providing enough fraud signal for model training. This hybrid nature has one important implication: model performance on this dataset may overestimate real-world performance, since synthetic fraud cases are by definition more "textbook" than the fraud patterns encountered in practice.

---

Quick Facts

Records	10,000
Columns	31 (27 used in modelling)
Fraud rate	10% (1,000 fraud cases)
Target variable	fraud_flag
Year covered	2023

---

Target Variable

Column	Type	Description
fraud_flag	int (0/1)	Target. 1 = fraudulent return, 0 = legitimate. 10% of records are fraud.

---

Taxpayer Profile

Column	Type	Description
taxpayer_id	string	Unique identifier (e.g. TP006253). No predictive value — dropped before modelling.
age	int	Taxpayer age. Range: 18–84, mean ≈ 51.
filing_status	string	Marital/household status: Single, Married, Head of Household.
occupation_category	string	Broad job category: Management, Professional, Sales, Service, Manual Labor, Administrative, Technical, Self-employed.
industry_sector	string	Industry the taxpayer works in: Healthcare, Finance, Manufacturing, Transportation, Construction, Retail, Education, Technology, Hospitality, Other.
geographic_region	string	US region: Northeast, Southeast, Midwest, Southwest, West.
years_filing	int	How many years the taxpayer has been filing returns. Range: 1–39.
num_dependents	int	Number of dependents claimed. Range: 0–5.
self_employed	int (0/1)	Whether the taxpayer is self-employed.
foreign_income	int (0/1)	Whether the taxpayer reported any foreign income.
rental_income	int (0/1)	Whether the taxpayer reported rental income.
previous_audit_flag	int (0/1)	Whether the taxpayer has been audited before. ~10% of records.

---

Financial Figures

Column	Type	Description
income_reported	float	Total income reported (USD). Range: ~$3,500–$721,000, mean ≈ $74,900.
investment_income	float	Income from investments. Mean ≈ $7,671.
business_expenses	float	Business-related expenses claimed. Many records show $0 (non-self-employed filers).
deductions_claimed	float	Total deductions claimed. Mean ≈ $24,350.
charitable_contributions	float	Charitable donation deductions. Mean ≈ $4,022.
medical_expenses	float	Medical expense deductions. Mean ≈ $1,978.
education_expenses	float	Education-related deductions. Mean ≈ $1,982.
mortgage_interest	float	Mortgage interest deduction. Mean ≈ $8,022.
property_tax	float	Property tax deduction. Mean ≈ $1,988.
state_local_tax	float	State and local tax deduction. Mean ≈ $4,570.
tax_credits_claimed	float	Total tax credits claimed. Mean ≈ $5,024.

---

Filing Details

Column	Type	Description
filing_date	string (date)	Date the return was filed. All records are from 2023. Converted to filing_month before modelling.
filing_method	string	How the return was submitted: Electronic or Paper.
days_to_deadline	int	Days between filing date and the tax deadline. Negative values indicate late filing. Range: -1 to 104.

---

Derived Ratio Features (dropped before modelling)

These columns were computed directly from other columns already present in the dataset. Keeping them would introduce data leakage — the model could rely on the pre-computed ratios rather than learning from the underlying raw features, which would overstate performance.

Column	Formula
deduction_to_income_ratio	deductions_claimed / income_reported
credit_to_income_ratio	tax_credits_claimed / income_reported
expense_per_dependent	deductions_claimed / num_dependents
income_per_dependent	income_reported / num_dependents

---

Preprocessing Summary

The following steps were applied before model training (see data_prep.ipynb):

1. Dropped taxpayer_id — unique identifier with no predictive signal.
2. Dropped the four derived ratio columns listed above — data leakage risk.
3. Extracted filing_month from filing_date, then dropped filing_date — all records are from 2023, so the year carries no variance.
4. Encoded categorical columns to integers: manual mapping for filing_status, filing_method, and geographic_region; dynamic mapping for occupation_category and industry_sector.
5. Exported cleaned data to synthetic_tax_fraud_dataset_preprocessed.csv.

Note: the dynamic encoding for occupation_category and industry_sector depends on unique() ordering, which is not guaranteed to be consistent across runs. For a production pipeline, a fixed mapping or a saved LabelEncoder would be more robust.

---

Modelling Notes

• Primary metric: PR-AUC and F1-score. With a 10% fraud rate, accuracy is misleading — a model that always predicts "no fraud" would still score 90%. Precision-Recall AUC is more informative under class imbalance.
• Class imbalance handling was explored via three strategies: no balancing (baseline), SMOTE + undersampling to 20% and 33% fraud ratios, and class_weight='balanced'.
• Threshold tuning was performed on the validation set — the default 0.5 threshold is rarely optimal on imbalanced data.
• Train / validation / test split was done with stratification to preserve the fraud rate across all subsets. The test set was held out until final evaluation.