lesson3.2:Unsupervised Learning, Anomaly Detection, and Market Structure Discovery in Financial Networks

Learning Objectives:

Master Isolation Forests by understanding recursive partitioning, path length distributions, and anomaly scoring for global outlier detection.
Apply Local Outlier Factor (LOF) to detect local density anomalies, crucial for identifying spoofing and quote stuffing in high-frequency markets.
Design and train Deep Autoencoders to learn compressed latent representations of normal behavior and detect novel fraud via reconstruction error.
Model financial ecosystems as graphs using Graph Neural Networks (GNNs) and message passing to uncover money laundering rings and hidden beneficial ownership.
Leverage Uniform Manifold Approximation and Projection (UMAP) for nonlinear dimensionality reduction to visualize financial contagion, regime shifts, and systemic risk.

Part 1: Isolation Forests and Density-Based Anomaly Detection

Traditional clustering algorithms such as K-Means attempt to identify dense groups of similar observations. However, financial fraud is extremely rare. Instead of finding clusters, institutional fraud systems focus on identifying isolated observations.

1.1: Why Isolation Beats Clustering for Fraud

Fraudulent transactions are numerically scarce and intentionally disguised to blend in. Clustering algorithms often fail because they force anomalies into existing clusters. Isolation algorithms explicitly exploit the key property of anomalies: they are few and different, meaning they are easier to isolate.

1.2: Isolation Forest – Recursive Partitioning

Isolation Forest is an ensemble learning algorithm specifically designed for anomaly detection. Unlike distance-based algorithms, Isolation Forest isolates observations through random recursive partitioning.

Recursive Partitioning (Isolation Process):
┌─────────────────────────────────────────────────────────────────────┐
| For each tree in the forest:                                     |
|  1. Randomly select one feature.                                 |
|  2. Randomly choose a split value within that feature's range.   |
|  3. Partition the data into two branches.                        |
|  4. Repeat recursively until every observation is isolated.      |
|                                                                  |
| Visualizing Path Lengths:                                        |
|                                                                  |
|   Normal Observation (Long Path):                                |
|   Root ──── Split 1 ──── Split 2 ──── Split 3 ──── Leaf         |
|   (Takes many random cuts to separate)                          |
|                                                                  |
|   Anomaly (Short Path):                                          |
|   Root ──── Split 1 ──── Leaf                                   |
|   (Isolated after very few cuts, due to extreme feature values) |
└─────────────────────────────────────────────────────────────────────┘

Path Length: For an observation x, let h(x) represent the number of splits required to isolate it.

Normal observations: Long path lengths (difficult to isolate).
Anomalies: Short path lengths (easy to isolate).

1.3: Anomaly Score Normalization

Across an ensemble of trees, the average path length is E[h(x)]. Isolation Forest normalizes these path lengths using a reference constant c(n) derived from unsuccessful searches in Binary Search Trees (BSTs):

Normalization Constant:
c(n) = 2H(n − 1) − 2(n − 1)/n

Where H(n) ≈ ln(n) + 0.5772156649 (Euler-Mascheroni constant)

The final Anomaly Score is:

s(x, n) = 2^( − E[h(x)] / c(n) )

Interpretation:
┌─────────────────────────────────────────────────────────────────────┐
| s(x,n) → 1.0  : Highly anomalous (isolated very quickly).        |
| s(x,n) ≈ 0.5  : Indistinguishable from normal observations.      |
| s(x,n) << 0.5 : Almost certainly a normal observation.           |
└─────────────────────────────────────────────────────────────────────┘

Financial Applications: Credit card fraud detection, insider trading surveillance, market abuse monitoring, and operational risk management.

1.4: Local Outlier Factor (LOF) – Local Density Anomalies

While Isolation Forest evaluates anomalies globally, Local Outlier Factor (LOF) compares an observation with its immediate neighborhood. This makes LOF indispensable for detecting local anomalies in datasets with varying densities (e.g., detecting spoofing orders in thinly traded vs. highly liquid stocks).

Local Reachability Density (LRD) of an observation x is:

LRD_k(x) = 1 / [ (1 / |N_k(x)|) Σ_(o ∈ N_k(x)) reach-dist_k(x,o) ]

Where Nₖ(x) are the k nearest neighbours, and reach-dist is the maximum of the actual distance and the distance to the k-th neighbour.

The LOF Score is the average ratio of the local reachability density of a point to that of its neighbours:

LOF_k(x) = (1 / |N_k(x)|) Σ_(o ∈ N_k(x)) LRD_k(o) / LRD_k(x)

Interpretation:
┌─────────────────────────────────────────────────────────────────────┐
| LOF ≈ 1 : Observation has similar density to its neighbors.      |
| LOF > 1 : Observation lies in a lower-density region (Anomaly).  |
| Large LOF values indicate significant anomalies (e.g., spoofing). |
└─────────────────────────────────────────────────────────────────────┘

Part 2: Deep Autoencoders for Fraud Detection

When financial data contains thousands of correlated variables, linear methods like PCA become insufficient. Deep Autoencoders learn compact, nonlinear latent representations capable of modeling the full complexity of normal transactional behaviour.

2.1: The Hourglass Architecture

An autoencoder consists of two neural networks:

Deep Autoencoder Architecture:
┌─────────────────────────────────────────────────────────────────────┐
|  Input Layer (x ∈ ℝⁿ)                                            |
|  ┌─────────────────────────────────────────────────────────────┐   |
|  │  │                                                         │   │
|  │  ▼                                                         │   |
|  │  ┌─────────────────────────────────────────────────────┐   │   |
|  │  │  Encoder: Compresses the input into a bottleneck   │   │   |
|  │  │  z = f_θ(x) = σ(Wₓx + bₓ)                        │   │   |
|  │  │  (Dimensionality reduces: n → d, where d << n)    │   │   |
|  │  └─────────────────────────────────────────────────────┘   │   |
|  │  │                                                         │   |
|  │  ▼  (Bottleneck Latent Space z ∈ ℝᵈ)                     │   |
|  │  ┌─────────────────────────────────────────────────────┐   │   |
|  │  │  Decoder: Reconstructs the input from the latent   │   │   |
|  │  │  x̂ = g_φ(z) = σ(W_zz + b_z)                       │   │   |
|  │  │  (Dimensionality expands: d → n)                   │   │   |
|  │  └─────────────────────────────────────────────────────┘   │   |
|  │  │                                                         │   |
|  │  ▼                                                         │   |
|  │  Output Layer (x̂ ∈ ℝⁿ)                                    │   |
|  └─────────────────────────────────────────────────────────────┘   |
└─────────────────────────────────────────────────────────────────────┘

2.2: Reconstruction Loss and Anomaly Threshold

Training minimizes the reconstruction error, typically using Mean Squared Error (MSE):

L(x, x̂) = ||x − x̂||²₂ = Σ[i=1→n] (x_i − x̂_i)²

Anomaly Detection: A transaction is flagged as anomalous if the reconstruction error E = ||x - x̂||²₂ exceeds a pre-defined threshold τ (e.g., the 99th percentile of reconstruction errors on a clean validation set).

Financial Applications:

Detecting previously unseen fraud patterns (zero-day attacks).
Identifying synthetic identities and account takeovers.
Cybersecurity intrusion detection in core banking systems.

Part 3: Graph Neural Networks (GNNs) for Anti-Money Laundering

Financial crime rarely involves isolated individuals. Instead, criminal organizations create interconnected networks of shell companies, mule accounts, and cryptocurrency wallets. These relationships are naturally represented as graphs.

3.1: Financial Graph Representation

A financial ecosystem is represented as a graph G = (V, E), where:

V (Vertices): Entities such as bank accounts, individuals, companies, or wallets.
E (Edges): Transactions between entities, containing attributes like amount, currency, timestamp, and location.

Financial Transaction Graph:
┌─────────────────────────────────────────────────────────────────────┐
|                                                                  |
|    [Cayman Corp] ───────$10M──────▶ [Crypto Exchange]           |
|          │                           │                           |
|          │ $5M                       │ $9.5M                     |
|          ▼                           ▼                           |
|    [Shell A] ──────▶ [Mule 1] ──────▶ [Offshore Bank]           |
|          ▲                           │                           |
|          │ $2M                       │ $8M                       |
|          │                           ▼                           |
|    [Trader A] ──────────────▶ [Mule 2]                           |
|                                                                  |
|  (Illicit funds are obscured through layered transactions)      |
└─────────────────────────────────────────────────────────────────────┘

3.2: Graph Convolutional Networks (GCNs) – Message Passing

Unlike traditional ML, Graph Neural Networks learn from both node features AND graph structure using message passing.

The node embedding update rule for a GCN layer is:

h_v^(l+1) = σ( W^(l) · Σ[u ∈ N(v) ∪ {v}] ( 1 / c_vu ) h_u^(l) )

Where:

hᵥ^(l) is the embedding of node v at layer l.
N(v) is the set of neighbours of node v.
W^(l) is the learnable weight matrix.
c_vu is a normalization constant (e.g., product of square root of degrees).

Interpretation: Each account updates its representation by aggregating information from its neighbouring accounts (e.g., who sent money, who received it). After several layers, the model learns that a shell company and its mule accounts effectively form a single “super-node” representing the fraud ring.

Financial Applications:

Anti-Money Laundering (AML): Identifying hidden beneficial ownership.
Fraud ring detection and terrorist financing surveillance.
Cryptocurrency forensic analysis.

Part 4: Manifold Learning and Financial Contagion Mapping

Financial markets often exhibit highly nonlinear relationships. During crises, asset correlations change dramatically, causing traditional linear methods like PCA to fail.

4.1: UMAP – Uniform Manifold Approximation and Projection

UMAP is a nonlinear dimensionality reduction algorithm grounded in Riemannian geometry and algebraic topology. Unlike PCA, UMAP attempts to preserve the topological structure (both local and global) of the data.

Step 1: High-Dimensional Graph Construction
UMAP constructs a weighted nearest-neighbour graph. The edge probability between observations i and j is:

p_ij = exp( − ( max(0, d(x_i,x_j) − ρ_i) ) / σ_i )

Where:

d(x_i,x_j) is the distance between observations.
ρ_i is the distance to the nearest neighbour (local connectivity).
σ_i is a scaling parameter.

Step 2: Low-Dimensional Optimization
UMAP learns a low-dimensional embedding Y ∈ ℝ² by minimizing the cross-entropy between the high-dimensional graph and the low-dimensional graph:

C = Σ_(i≠j) [ p_ij log(p_ij / q_ij) + (1 − p_ij) log((1 − p_ij)/(1 − q_ij)) ]

Where q_ij is the similarity in the low-dimensional embedding.

UMAP Topological Preservation:
┌─────────────────────────────────────────────────────────────────────┐
|  High-Dimensional Financial Data (Hundreds of indicators):       |
|  ┌─────────────────────────────────────────────────────────────┐   |
|  │  (Dense clusters of correlated assets)                     │   |
|  │  ╲ ╱ ╲ ╱                                                   │   |
|  │   ╲╱  ╲╱                                                    │   |
|  │   ●───●───●  (Global market structure)                     │   |
|  └─────────────────────────────────────────────────────────────┘   |
|                              ▼  (UMAP Projection)                 |
|  Low-Dimensional Visualization (2D/3D):                          |
|  ┌─────────────────────────────────────────────────────────────┐   |
|  │  ●●●●●●●●                                                    │   |
|  │  ●   ●   ●                                                   │   |
|  │  ●  ╲ ╱  ●  (Regime 1: Risk-On)                            │   |
|  │  ●   ●   ●                                                   │   |
|  │      ▲                                                       │   |
|  │      │ (Contagion Path)                                      │   |
|  │      ▼                                                       │   |
|  │  ○○○○○○○○                                                    │   |
|  │  ○   ○   ○  (Regime 2: Crisis / Risk-Off)                  │   |
|  └─────────────────────────────────────────────────────────────┘   |
└─────────────────────────────────────────────────────────────────────┘

Financial Applications:

Visualizing systemic financial contagion and sovereign debt crises.
Mapping banking sector interconnectedness.
Detecting market regime transitions (Bull, Bear, Crisis).
Portfolio clustering and diversification analysis.

Practical Implementation Playbook (Python)

Below is a production-inspired pipeline implementing Isolation Forest, a Deep Autoencoder, and a Graph Neural Network (GCN) for AML.

import numpy as np
import torch
import torch.nn as nn
from sklearn.ensemble import IsolationForest
from sklearn.neighbors import LocalOutlierFactor
from torch_geometric.nn import GCNConv
from torch_geometric.data import Data

# -------------------- 1. GLOBAL ANOMALY DETECTION --------------------
# Isolation Forest for transaction fraud
model_iso = IsolationForest(
    n_estimators=100, 
    contamination=0.001,  # Expected fraud rate
    random_state=42
)
# decision_function gives the anomaly score (higher = more anomalous)
scores = model_iso.fit(X_train).decision_function(X_test)
anomalies = np.where(scores < 0, 1, 0)  # Negative scores indicate outliers

# -------------------- 2. LOCAL ANOMALY DETECTION --------------------
# LOF for detecting spoofing in order book data
model_lof = LocalOutlierFactor(
    n_neighbors=20, 
    contamination=0.01, 
    novelty=True
)
model_lof.fit(X_train)
lof_scores = -model_lof.score_samples(X_test)  # Higher scores = more anomalous

# -------------------- 3. DEEP AUTOENCODER (PyTorch) --------------------
class DeepAutoencoder(nn.Module):
    def __init__(self, n_features, latent_dim=8):
        super().__init__()
        # Encoder
        self.encoder = nn.Sequential(
            nn.Linear(n_features, 64),
            nn.ReLU(),
            nn.Linear(64, 16),
            nn.ReLU(),
            nn.Linear(16, latent_dim)
        )
        # Decoder
        self.decoder = nn.Sequential(
            nn.Linear(latent_dim, 16),
            nn.ReLU(),
            nn.Linear(16, 64),
            nn.ReLU(),
            nn.Linear(64, n_features),
            nn.Sigmoid()
        )
    
    def forward(self, x):
        z = self.encoder(x)
        return self.decoder(z)

# Train using MSE loss; threshold set at 95th percentile of reconstruction error
model_ae = DeepAutoencoder(n_features=100)
# ... training loop omitted for brevity ...
recon_error = torch.mean((x_test - model_ae(x_test))**2, dim=1)
anomalies = recon_error > threshold

# -------------------- 4. GRAPH NEURAL NETWORK (PyTorch Geometric) --------
class AML_GCN(nn.Module):
    def __init__(self, in_channels, hidden_channels, out_channels=2):
        super().__init__()
        self.conv1 = GCNConv(in_channels, hidden_channels)
        self.conv2 = GCNConv(hidden_channels, out_channels)
    
    def forward(self, x, edge_index):
        # x: Node features (e.g., avg transaction size, velocity)
        # edge_index: Transaction edges (sender, receiver)
        x = self.conv1(x, edge_index).relu()
        x = self.conv2(x, edge_index)
        return x  # Logits for AML classification (if labels exist) or embeddings

# Construct graph data (2 nodes, 1 transaction)
edge_index = torch.tensor([[0, 1], [1, 0]], dtype=torch.long)
x = torch.tensor([[1.2, 0.5], [0.8, 0.7]], dtype=torch.float)  # Features
data = Data(x=x, edge_index=edge_index)

Summary

Advanced unsupervised learning enables financial institutions to detect hidden structures, emerging risks, and previously unseen forms of financial crime without relying on historical labels.

Isolation Forests detect anomalies by recursively partitioning data and measuring the expected path length required to isolate each observation—shorter path lengths correspond to more anomalous behaviour. Local Outlier Factor (LOF) extends this by identifying local density anomalies, making it indispensable for detecting market manipulation techniques such as spoofing and quote stuffing in micro-structure data.

Deep Autoencoders learn compressed latent representations of normal financial behaviour and identify anomalies through reconstruction error, enabling the discovery of novel fraud patterns that rule-based systems miss. Graph Neural Networks (GNNs) model financial ecosystems as interconnected graphs and use message passing to uncover hidden fraud rings, money laundering networks, and beneficial ownership structures that are invisible in tabular datasets.

Finally, Uniform Manifold Approximation and Projection (UMAP) performs nonlinear dimensionality reduction while preserving topological relationships, allowing analysts to intuitively visualize financial contagion, systemic risk propagation, and evolving market regimes.

Together, these techniques form the foundation of modern institutional anomaly detection, anti-money laundering systems, financial crime analytics, and systemic risk monitoring, enabling quantitative analysts to uncover patterns that traditional supervised and linear methods cannot detect.

Key Terminology Glossary

Term	Definition
Isolation Path Length	The number of recursive splits required to completely isolate an observation in an Isolation Tree.
Anomaly Score (s)	A normalized score derived from the average isolation path length; values near 1 indicate strong anomalies.
Local Reachability Density (LRD)	The inverse of the average reachability distance to the `k` nearest neighbours of a point.
Local Outlier Factor (LOF)	A score comparing the LRD of a point to the average LRD of its neighbours; values >> 1 indicate anomalies.
Bottleneck (Latent Space)	The low-dimensional layer in an autoencoder that forces the network to learn the most compressed representation of the input data.
Reconstruction Error	The difference between the original input and its decoded output; used as an anomaly metric.
Message Passing	The core mechanism of GNNs where nodes iteratively aggregate feature information from their neighbours.
Topological Preservation	UMAP’s ability to retain both the local neighbourhood structure and global density patterns of high-dimensional data when projecting to 2D/3D.
Cross-Entropy (UMAP)	The loss function UMAP minimizes to align the low-dimensional embedding with the high-dimensional topological representation.
Contagion Mapping	Visualizing how shocks propagate through interconnected financial networks, often using UMAP to reveal crisis pathways.

Certificate in Artificial Intelligence for Finance lesson3.2:Unsupervised Learning, Anomaly Detection, and Market Structure Discovery in Financial Networks

Certificate in Artificial Intelligence for Finance

Learning Objectives:

Part 1: Isolation Forests and Density-Based Anomaly Detection

1.1: Why Isolation Beats Clustering for Fraud

1.2: Isolation Forest – Recursive Partitioning

1.3: Anomaly Score Normalization

1.4: Local Outlier Factor (LOF) – Local Density Anomalies

Part 2: Deep Autoencoders for Fraud Detection

2.1: The Hourglass Architecture

2.2: Reconstruction Loss and Anomaly Threshold

Part 3: Graph Neural Networks (GNNs) for Anti-Money Laundering

3.1: Financial Graph Representation

3.2: Graph Convolutional Networks (GCNs) – Message Passing

Part 4: Manifold Learning and Financial Contagion Mapping

4.1: UMAP – Uniform Manifold Approximation and Projection

Practical Implementation Playbook (Python)

Summary

Key Terminology Glossary

International BSF

START YOUR LEARNING JOURNEY