README.md

---
license: cc-by-4.0
task_categories:
  - tabular-classification
  - tabular-regression
  - time-series-forecasting
language:
  - en
tags:
  - synthetic
  - finance
  - fraud-detection
  - aml
  - anti-money-laundering
  - anomaly-detection
  - transaction-monitoring
  - fintech
  - banking
  - crypto
  - market-manipulation
  - account-takeover
  - false-positive-reduction
pretty_name: Aestrea Dark Pool Financial Anomaly Pack
size_categories:
  - 10K<n<100K
configs:
  - config_name: default
    data_files:
      - split: train
        path: dark_pool_sample.parquet
---

# Aestrea Dark Pool Financial Anomaly Pack (Sample)

**A synthetic financial-anomaly and fraud-detection dataset for AML / transaction-monitoring model training, false-positive reduction research, and fintech anomaly-detection pipelines.** Each row is a complete financial activity sequence — a dormant warm-up period followed by a sudden activity burst — labeled as account takeover, market-manipulation spoofing, or legitimate-but-high-risk (false positive) behavior, with telemetry, identity context, detection logic, and financial exposure.

Built by [SolsticeAI](https://www.solsticestudio.ai/datasets) as a free sample of a larger commercial pack. 100% synthetic. No real transactions, accounts, addresses, order books, or customer data.

## What is included

| File | Rows | Format | Purpose |
|---|---:|---|---|
| `dark_pool_sample.parquet` | 10,000 | Parquet | Columnar, typed, best for analytics |
| `dark_pool_sample.jsonl` | 10,000 | JSON Lines | Streaming / LLM training friendly |

**This sample:** 10,000 financial activity sequences, stratified 3,333 per fraud class.  
**Fraud classes (3):** `Account_Takeover_Sequence`, `Market_Manipulation_Spoofing`, `Legitimate_High_Risk_Activity` (false positive)  
**Label distribution:** ~6,666 `fraudulent` / ~3,334 `benign` (by design — one of three scenarios is a benign high-risk lookalike)  
**Severity tiers:** `low`, `medium`, `high`, `critical` — scenario-weighted so spoofing skews mid, ATO skews high, false-positives skew low  
**Assets covered:** BTC, ETH, USDC, USDT, SOL, USD, EUR, GBP, JPY  
**Geographic coverage:** 20 country codes across high-risk and low-risk jurisdictions

## Record structure

Each record is one financial activity sequence with 7 top-level fields:

| Field | Type | Contents |
|---|---|---|
| `schema_version` | string | Pack schema version (`1.0.0-dark-pool-sample`) |
| `event` | struct | `id`, `trace_id`, `timestamp`, `severity`, `label`, `label_confidence` |
| `identity_context` | struct | `account_id`, `baseline_risk_score`, `session_entropy`, `account_age_days`, `kyc_tier` |
| `telemetry_stream` | list<struct> | Ordered financial events: `timestamp`, `event_name` (e.g., `LARGE_WITHDRAWAL`, `LIMIT_ORDER_CANCELLED`, `LOGIN_NEW_COUNTRY`), `asset`, `transaction_amount_usd`, `geo_country`, `latency_ms`, `burst_indicator` |
| `vulnerability` | struct | `fraud_class`, `scenario_description`, `exposure_vector` |
| `detection` | struct | `anomaly_type`, `baseline_deviation`, `detection_logic` (SQL-like), `anomaly_score`, `confidence_band` |
| `impact` | struct | `financial_exposure_usd`, `customer_funds_at_risk_usd`, `recoverable_pct` |
| `simulation` | struct | `synthetic`, `engine`, `chaos_profile`, `ground_truth_classification` |

See [SCHEMA.md](./SCHEMA.md) for the full nested field breakdown.

## Why this dataset is useful

Most public financial-fraud datasets are either flat tabular snapshots (IEEE-CIS, credit-card fraud) or narrow single-event labels. Real AML / transaction-monitoring models need something these don't provide: **sequences** (warm-up → burst), **ambiguous labels** (legitimate-but-high-risk that looks like fraud), and **detection logic grounded in behavioral patterns** rather than just feature engineering.

- Full pre-event dormant and check-in telemetry — not just the burst
- `burst_indicator` flags at each step so models can learn temporal dynamics explicitly
- A dedicated false-positive class (`Legitimate_High_Risk_Activity`) with benign ground truth but fraud-adjacent behavior — the exact class that drives real-world FP rates
- Per-record `detection_logic` in SQL-like form shows how the scenario would be detected in production
- Per-record `anomaly_score` and `label_confidence` so models can train on calibration targets, not just hard labels
- Cross-asset (crypto + fiat), cross-geography coverage

## Typical use cases

- Fraud-detection model training
- AML / transaction-monitoring pipelines
- False-positive-reduction research (reducing customer-experience pain)
- Risk-scoring systems
- Account-takeover detection
- Spoofing / market-manipulation classifiers
- Temporal-burst anomaly detection
- LLM fine-tuning on fraud-investigation narratives
- Benchmarking anomaly scoring against ambiguous ground truth

## Quick start

```python
import pandas as pd
import pyarrow.parquet as pq

df = pq.read_table("dark_pool_sample.parquet").to_pandas()

# Label distribution (mix of fraudulent and benign)
print(df["event"].apply(lambda e: e["label"]).value_counts())

# Average financial exposure by fraud class
df["cls"] = df["vulnerability"].apply(lambda v: v["fraud_class"])
df["exposure"] = df["impact"].apply(lambda i: i["financial_exposure_usd"])
print(df.groupby("cls")["exposure"].mean().round(0))

# Burst density by scenario (bursts per sequence)
df["burst_count"] = df["telemetry_stream"].apply(
    lambda s: sum(1 for step in s if step["burst_indicator"])
)
print(df.groupby("cls")["burst_count"].mean().round(2))

# False-positive rate signal — proportion of fraudulent-label vs benign per severity
df["severity"] = df["event"].apply(lambda e: e["severity"])
df["label"] = df["event"].apply(lambda e: e["label"])
print(pd.crosstab(df["severity"], df["label"], normalize="index").round(3))
```

Streaming form:

```python
import json

with open("dark_pool_sample.jsonl") as f:
    for line in f:
        sequence = json.loads(line)
        # one financial activity sequence per line
```

## Responsible use

This dataset is intended for **defensive / monitoring** use cases: fraud-detection model training, AML research, false-positive reduction, and academic benchmarks. It contains synthesized transaction amounts, synthetic account IDs, and fictional scenario narratives — it does **not** contain real customer data, real transactions, real wallet addresses, or any PII. Models trained on this data will learn fraud-pattern shape and temporal dynamics; downstream deployment for live transaction monitoring requires calibration against institution-specific real-world data under appropriate compliance review.

## License

Released under **CC BY 4.0**. Use freely for research, fraud-tool prototyping, education, and commercial development with attribution.

## Get the full pack

This Hugging Face repo is a **10K-sequence sample**. The production pack scales to 5M+ sequences with expanded scenario coverage (layering, smurfing, wash trading, sanctions-evasion routing, trade-based ML, bust-out fraud), additional asset classes (equities, FX, derivatives, NFTs), multi-entity transaction graphs, richer ambiguous-label distribution tuning, parquet + JSONL + SAR-narrative-aligned delivery, and buyer-specific variants.

**Self-serve (Stripe checkout):**
- [**Sample Scale tier — $5,000**](https://buy.stripe.com/7sY5kD2j85QTfSb5lfeEo03) — ~25K records, one subject, 72-hour delivery.

**Full pack + enterprise scope:**
- [www.solsticestudio.ai/datasets](https://www.solsticestudio.ai/datasets) — per-SKU pricing across Starter / Professional / Enterprise tiers, plus commercial licensing, custom generation, and buyer-specific variants.

**Procurement catalog:**
- [SolsticeAI Data Storefront](https://solsticeai.mydatastorefront.com) — available via Datarade / Monda.

## Citation

```bibtex
@dataset{solstice_dark_pool_pack_2026,
  title        = {Aestrea Dark Pool Financial Anomaly Pack (Sample)},
  author       = {SolsticeAI},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/solsticestudioai/dark-pool-pack}
}
```