Forge a Mobile Money fraud-signal dataset
Choose how many transaction records you want and a seed. The same two numbers will always produce the same dataset — write the seed down and cite it in your methodology chapter.
Advanced: customize the model
Same directional relationships described in the methodology — strengthen, weaken, or rebalance them here without touching any code. Defaults are illustrative, anchored against publicly-documented Mobile Money fraud typologies, not derived from real transaction data.
Fraud risk (frequency)
Every factor below feeds one logistic model over is_fraudulent. Raise or lower a loading to strengthen or weaken that hypothesis.
Loss given fraud (severity) two-stage only
Cite this dataset
This is a specific, reproducible dataset — seed, record count, generator version, and any advanced-config changes are all part of what makes it citable. Copy a citation below, or download the full provenance manifest for a supervisor or reviewer to verify against.
This is synthetic data, not real records — safe to use freely for research, coursework, or analysis, but it should never be presented as real-world data. If this dataset (or results derived from it) appears in published, submitted, or graded work, please cite the specific seed and generator version using the citation tool below.
Also downloadable as from your account's Generation history.
Data dictionary (0 columns)
| Column | Type | Unit | Description |
|---|
Hypothesis check
Directional confirmation on this batch — sign and rough magnitude should match, exact numbers will vary run to run except when the seed is held fixed. Four hypotheses act on fraud probability directly and hold in both stage modes; two more act on loss given fraud for transactions already flagged fraudulent, and only exist — and are only shown — in two-stage mode (see the methodology page). Agent cash-out channel appears on both sides, reading two different things off the same field. Every card scoped to Cash-Out and/or fraudulent rows needs a larger batch than the others to look stable — the frequency-side agent-channel card needs at least 20,000 records; the severity-side one needs considerably more (at least 200,000) since it's scoped to Cash-Out and fraudulent rows together.