Transaction & Payment Processing Data: Labeled Fraud Benchmarks, National Rails and Adoption Panels · Head-to-head
IEEE-CIS Fraud Detection (Vesta) vs RBI Payment System Indicators: Labeled Transactions or Official System Totals
Which transaction & payment processing data: labeled fraud benchmarks, national rails and adoption panels data fits your job: Kaggle – IEEE-CIS Fraud Detection – 590K+ Transaction and Identity Tables, or Reserve Bank of India – Payment System Indicators. API, files, or your warehouse. Daily, weekly, or hourly.
Kaggle – IEEE-CIS Fraud Detection – 590K+ Transaction and Identity Tables (Vesta)
Reserve Bank of India – Payment System Indicators (Monthly Volume and Value)
Where the fields line up
No shared field names. These two answer different questions.
| Field | Kaggle – IEEE-CIS Fraud Detection – 590K+ Transaction and Identity Tables | Reserve Bank of India – Payment System Indicators |
|---|---|---|
TransactionID | Primary key joining the transaction and identity tables. | not in this set |
isFraud | Target variable: 1 for fraudulent transaction, 0 otherwise (present in training data only). | not in this set |
TransactionDT | Timedelta in seconds from a reference datetime; not an actual wall-clock timestamp. | not in this set |
TransactionAmt | Payment amount in USD-equivalent masked units. | not in this set |
ProductCD | Product code for each purchase (e.g. W, C, R, H, S). | not in this set |
card1-card6 | Payment card metadata: card number prefix, issuing bank, card network/type and country. | not in this set |
addr1, addr2 | Masked billing address region and country codes. | not in this set |
dist1, dist2 | Distance measures between billing/shipping addresses or card/IP geographies. | not in this set |
P_emaildomain, R_emaildomain | Purchaser and recipient email domains (e.g. gmail.com); recipient domain often null except for certain product types. | not in this set |
C1-C14 | Counting features such as how many addresses, cards or phone numbers are associated with the payment instrument. | not in this set |
D1-D15 | Timedelta features such as days since previous transaction or since card first use. | not in this set |
M1-M9 | Match flags (T/F/NaN) indicating whether names, addresses or amounts match across transaction parties. | not in this set |
Coverage, side by side
| Kaggle – IEEE-CIS Fraud Detection – 590K+ Transaction and Identity Tables | Reserve Bank of India – Payment System Indicators | |
|---|---|---|
| Geographic | Global e-commerce transactions, with countries masked as numeric codes and no merchant identifiers | India, domestic transactions only; Part IV adds international use of India-issued cards and prepaid instruments |
| Temporal | About six months of transactions split chronologically into train and test periods via the TransactionDT timedelta | Monthly editions archived from 1990; current five-part format with granular mode detail from November 2019; Part V fraud series from September 2022 |
| Granularity | One row per online payment transaction, optionally enriched with device and identity attributes | Monthly totals per payment system or mode, fiscal-year-to-date aggregates, and month-end infrastructure stock counts |
What each contains
They tie on 1 attribute. Pick by fit, not by loyalty.
| Kaggle – IEEE-CIS Fraud Detection – 590K+ Transaction and Identity Tables | Reserve Bank of India – Payment System Indicators | |
|---|---|---|
| Geography | Global e-commerce transactions, with countries masked as numeric codes and no merchant identifiers | India, domestic transactions only; Part IV adds international use of India-issued cards and prepaid instruments |
| Temporal reach | About six months of transactions split chronologically into train and test periods via the TransactionDT timedelta | Monthly editions archived from 1990; current five-part format with granular mode detail from November 2019; Part V fraud series from September 2022 |
| Granularity | One row per online payment transaction, optionally enriched with device and identity attributes | Monthly totals per payment system or mode, fiscal-year-to-date aggregates, and month-end infrastructure stock counts |
| Scale | 590,540 rows x 394 transaction columns plus a 144,233-row x 41-column identity table, about 1.4 GB delivered | Roughly 100 data rows across Parts I-V per monthly edition, about 490 KB per edition |
| Best for | Training and benchmarking fraud models against a labeled target with realistic class imbalance | Sizing and tracking India's payment rails with official monthly volume, value, infrastructure and fraud statistics |
Or take both in one feed
Yes, as complements rather than a join - there is no shared key between a masked transaction ID and a national aggregate row. The productive pattern stacks them by role: train the detector on IEEE-CIS transaction-level patterns, then calibrate expectations with RBI's Part V ratios so projected loss rates reflect India's rails instead of a global e-commerce average. A team launching in India can quote 227,160.66 lakh June 2026 UPI transactions alongside a model validated on 590,540 labeled examples, and both numbers survive scrutiny.
Two practical notes from the records. Mind definitional breaks in the RBI series: card and prepaid figures from November 2019 onward are flagged as not comparable with earlier periods after revised definitions, and failed transactions, chargebacks, reversals and expired cards sit outside the counted totals. And treat the two fraud measures as different quantities - IEEE-CIS labels individual e-commerce transactions while Part V counts reported domestic frauds across all modes - so never divide one by the other. Or take both in one feed.
API, files, or your warehouse. Daily, weekly, or hourly.
Fair questions
Is the IEEE-CIS Fraud Detection dataset better than RBI Payment System Indicators?
Better depends on what you are building. IEEE-CIS wins on learning signal: 590,540 individual transactions with a labeled isFraud target at roughly a 3.5% positive rate plus 394 engineered features. RBI wins on market truth: official monthly volume and value for every major Indian rail. One trains models; one sizes markets.
Which dataset has the longer history?
The RBI release, decisively. Its archive reaches back to 1990, though the current five-part format with granular mode-level detail runs from November 2019 onward and the domestic fraud series is shown from September 2022. The IEEE-CIS benchmark covers about six months of transactions split chronologically into train and test periods - depth per row rather than length of calendar.
Do either of these cover card-present fraud at a specific merchant?
Neither. IEEE-CIS is online e-commerce transactions with countries masked to codes and no merchant identifiers, and it carries no point-of-sale detail beyond card metadata flags. RBI reports card payments in aggregate splits such as PoS versus others, plus terminal stock counts - never per-merchant records.
Can I use both datasets to estimate fraud rates for India?
Yes, as a bounded estimate rather than a merge. RBI's Part V publishes domestic payment frauds monthly - 3.53 lakh fraudulent transactions against total digital volume in its verified June 2026 edition, a fraud-to-payment-value ratio of 0.150 basis points. IEEE-CIS supplies the transaction-level feature patterns behind similar ratios. Join them on your own analysis, not on a shared key.
What does the identity table in IEEE-CIS add?
Device and network context for a subset of transactions: id_01 through id_38 covering browser versions, screen resolutions, IP-linked aggregates and operating systems, plus DeviceType and raw DeviceInfo strings. Only 144,233 of the 590,540 transactions carry identity records, so any model using both tables must handle the missing-device segment explicitly.
Are the field definitions documented for both datasets?
Both records in Datadory's catalog carry verified field dictionaries. IEEE-CIS documents every column family - TransactionDT, ProductCD, card1-card6, C1-C14 counts, D1-D15 timedeltas, M1-M9 match flags and the V1-V339 block whose semantics Vesta never disclosed. RBI's dictionary covers row labels, volume in lakh, value in rupee crores, fiscal-year columns and month-end stock counts.