aiacademy-kg/mbank-atm-bishkek

MBank ATM Dataset (Bishkek, 2024–2025)

Per-second dataset of a network of 120 ATMs over 731 days (2024-01-01 … 2025-12-31). Card-processing style records (ISO-8583-inspired), grounded in the real-world context of Kyrgyzstan: ERA5 weather, the USD/KGS rate of the National Bank of the KR, and ATM coordinates from OpenStreetMap.

Intended as a training ground: demand forecasting, ATM clustering, RFM analysis, anomaly detection. Every anomaly has a checkable "answer key" in labels/.

---

Google Colab — install & load (in the browser)

In a fresh Colab cell, install the libraries and download the dataset:

# 1. Libraries (pandas/pyarrow ship with Colab; only the HF client is needed)
!pip install -q huggingface_hub datasets

# 2. Download the whole dataset in one call
from huggingface_hub import snapshot_download
path = snapshot_download("aiacademy-kg/mbank-atm-bishkek", repo_type="dataset")
print("Dataset at:", path)

# 3. Read tables with plain pandas
import pandas as pd
dim_atm  = pd.read_parquet(f"{path}/dimensions/dim_atm.parquet")
calendar = pd.read_parquet(f"{path}/dimensions/dim_calendar.parquet")
fraud    = pd.read_parquet(f"{path}/labels/fraud.parquet")
print(dim_atm.shape, calendar.shape)

Transactions (12M rows) — read only the columns/months you need, not all at once:

import pyarrow.dataset as ds
tx = ds.dataset(f"{path}/facts/fact_transactions", partitioning="hive")
march = tx.to_table(
    filter=(ds.field("year") == 2024) & (ds.field("month") == 3),
    columns=["timestamp", "atm_id", "dispensed_amount", "response_code"],
).to_pandas()
print(march.shape)

Or via 🤗 datasets (each table is a separate subset):

from datasets import load_dataset
tx  = load_dataset("aiacademy-kg/mbank-atm-bishkek", "fact_transactions", split="train")
atm = load_dataset("aiacademy-kg/mbank-atm-bishkek", "dim_atm", split="train")
# subsets: fact_transactions, fact_cash_ops, fact_atm_status, dim_atm, dim_calendar,
#          dim_customer, dim_card, dim_district, labels_fraud, labels_cash_events, …

If the dataset is private, authenticate first: from huggingface_hub import login; login("hf_YOUR_TOKEN").

---

Quick start (local)

import pandas as pd
import pyarrow.dataset as ds

# Reference tables
dim_atm = pd.read_parquet("dataset/dimensions/dim_atm.parquet")
cal     = pd.read_parquet("dataset/dimensions/dim_calendar.parquet")

# Transaction fact — 12M rows, partitioned by year/month. Read columnar / filtered:
tx = ds.dataset("dataset/facts/fact_transactions", partitioning="hive")
march = tx.to_table(filter=(ds.field("year") == 2024) & (ds.field("month") == 3),
                    columns=["timestamp","atm_id","dispensed_amount","response_code"]).to_pandas()

# Anomaly ground truth
fraud = pd.read_parquet("dataset/labels/fraud.parquet")

⚠️ fact_transactions is 12M rows (~727 MB). Don't load it whole — read the columns you need and filter by the year=/month= partitions.

---

Folder layout

dataset/
├── README.md            ← this file
├── DATA_DICTIONARY.md   ← full field dictionary (types, values)
├── MANIFEST.md          ← auto-inventory: tables, row counts, sizes
├── dimensions/          ← reference tables (star schema)
│   ├── dim_district.parquet (+csv)   16 zones of Bishkek
│   ├── dim_calendar.parquet (+csv)   731 days: holidays, weather, FX rate
│   ├── dim_atm.parquet      (+csv)   120 ATMs (passport)
│   ├── dim_customer.parquet          550k customers
│   └── dim_card.parquet              700k cards (on-us)
├── facts/
│   ├── fact_transactions/            12M transactions, partitioned year=/month=
│   ├── fact_cash_ops.parquet         8.8k cash replenishments
│   └── fact_atm_status.parquet       9.8k status/downtime intervals
├── labels/              ← anomaly ground truth
│   ├── fraud.parquet          skimming + geo-impossibility
│   ├── cash_events.parquet    depletions (out_of_cash)
│   ├── faults.parquet         network faults / stuck terminals
│   ├── cold_start.parquet     new ATMs (truncated history)
│   ├── events.parquet         network-wide surges (FX shocks)
│   └── drift_spec.parquet     year-2 concept-drift parameters
└── external/            ← external real-world inputs (reference)
    ├── weather_daily.parquet  Open-Meteo ERA5 weather
    ├── fx_usd_kgs.parquet      NBKR exchange rate
    └── atm_locations.parquet   OSM coordinates

---

Star schema (table relationships)

                    ┌──────────────┐
                    │ dim_calendar │  date
                    └──────┬───────┘
                           │ date
   ┌────────────┐   ┌──────┴────────────┐   ┌──────────────┐
   │ dim_district│──│  fact_transactions │──│  dim_atm     │
   └────────────┘   │      (12M)        │   └──────────────┘
      district_id   └──────┬─────┬──────┘      atm_id
                    card_id│     │customer_id
                     ┌─────┴──┐ ┌┴───────────┐
                     │dim_card│ │dim_customer│
                     └────────┘ └────────────┘

Keys and relationships (JOIN):

From	Field	To	Note
fact_transactions	atm_id	dim_atm	ATM
fact_transactions	district_id	dim_district	zone (denormalized)
fact_transactions	date	dim_calendar	day (weekday/holiday/weather/FX)
fact_transactions	card_id	dim_card	card (on-us only; off-us → OFFUS-*, not in dim_card)
fact_transactions	customer_id	dim_customer	customer (null for off-us)
dim_card	customer_id	dim_customer	cardholder
fact_cash_ops / fact_atm_status	atm_id	dim_atm	ATM

on-us vs off-us: is_on_us=True — MBank card (present in dim_card/dim_customer). is_on_us=False — other bank / tourist: card_id like OFFUS-*, empty customer_id, an interchange fee_amount is charged.

---

Anomaly ground truth (labels/)

Every anomaly is parameterized and logged — each task has a checkable answer.

File	Anomaly	What to look for	Method
cash_events	Depletion 7.1	withdrawal drop + 99 spike	rules, DBSCAN over day-ATM
faults	Network fault 7.2 / stuck 7.8	91/92/96 runs, flatlined telemetry	downtime detection, variance
events	Surge 7.3	network-wide day anomaly (FX shock)	collective outlier
cold_start	New ATM 7.4	truncated history	model robustness
drift_spec	Concept drift 7.6	year-1 model degrades on year-2	drift monitoring
fraud	Skimming 7.7 / geo 7.10	night off-us runs, card at far ATMs	IsolationForest, velocity rules

fraud/cash_events reference concrete txn_id/atm_id+interval in the fact table, so labels can be matched exactly to transactions.

---

External inputs (external/)

The dataset is grounded in real-world context of Kyrgyzstan — these series actually drive behavior:

File	Source	Effect in the data
weather_daily	Open-Meteo ERA5 (CC BY 4.0)	frost/snow → drop at street ATMs
fx_usd_kgs	National Bank of the KR	rate jump → cash run
atm_locations	OpenStreetMap (ODbL)	ATM coordinates (placement skeleton)

Verified: frost on 16–17 Feb 2024 → street ATMs −22%; FX shock on 20 Mar 2025 (−1.41 KGS) → volume ×3.

Full field descriptions — in DATA_DICTIONARY.md.