Chapter 260: Knowledge Graph Trading

Knowledge graphs (KGs) encode entities and their relationships as a structured network, enabling machine learning models to reason over interconnected financial information. Unlike tabular features that treat each input independently, KGs capture cross-entity dependencies — such as supply-chain links between companies, sector memberships, executive board overlaps, and macroeconomic exposures — providing a richer context for trading decisions.

Key Concepts

Knowledge Graph Fundamentals

A knowledge graph $G = (V, E, R)$ consists of:

• Nodes $V$: entities (companies, sectors, commodities, people, events)
• Edges $E \subseteq V \times R \times V$: directed relationships between entities
• Relation types $R$: semantic labels (e.g., supplies_to, competes_with, member_of)

Each fact is stored as a triple $(h, r, t)$ where $h$ is the head entity, $r$ is the relation, and $t$ is the tail entity. For example: (AAPL, supplies_to, TSLA) or (NVDA, member_of, S&P500).

Knowledge Graph Embeddings

KG embedding models learn continuous vector representations of entities and relations by optimizing a scoring function $f(h, r, t)$ over observed triples.

TransE models relations as translations in embedding space:

$f(h, r, t) = -\|\mathbf{h} + \mathbf{r} - \mathbf{t}\|$

The loss encourages $\mathbf{h} + \mathbf{r} \approx \mathbf{t}$ for positive triples and $\mathbf{h} + \mathbf{r} \not\approx \mathbf{t}'$ for corrupted triples.

DistMult uses a bilinear scoring function:

$f(h, r, t) = \mathbf{h}^\top \text{diag}(\mathbf{r}) \, \mathbf{t} = \sum_i h_i \cdot r_i \cdot t_i$

ComplEx extends DistMult to complex-valued embeddings, handling asymmetric relations:

$f(h, r, t) = \text{Re}\left(\sum_i h_i \cdot r_i \cdot \bar{t}_i\right)$

Graph Neural Networks on KGs

Relational Graph Convolutional Networks (R-GCN) aggregate neighbor information through typed edges:

$\mathbf{h}_i^{(l+1)} = \sigma\left(\sum_{r \in R} \sum_{j \in \mathcal{N}_i^r} \frac{1}{c_{i,r}} \mathbf{W}_r^{(l)} \mathbf{h}_j^{(l)} + \mathbf{W}_0^{(l)} \mathbf{h}_i^{(l)}\right)$

where $\mathcal{N}_i^r$ is the set of neighbors of node $i$ under relation $r$, $c_{i,r}$ is a normalization constant, and $\mathbf{W}_r^{(l)}$ are relation-specific weight matrices.

Temporal Knowledge Graphs

Financial KGs evolve over time. Temporal KG models extend static embeddings with a time component:

$f(h, r, t, \tau) = \mathbf{h}_\tau^\top \text{diag}(\mathbf{r}_\tau) \, \mathbf{t}_\tau$

where entity and relation embeddings are time-dependent functions $\mathbf{h}_\tau = g(\mathbf{h}, \tau)$, capturing how corporate relationships, sector allocations, and supply chains change.

ML Approaches for Trading

Entity-Based Alpha Signals

KG embeddings provide entity features that complement traditional factor models. Given a stock’s embedding $\mathbf{e}_i$, we can train a classifier:

$P(\text{up}_i \mid \mathbf{e}_i, \mathbf{x}_i) = \sigma(\mathbf{w}^\top [\mathbf{e}_i \| \mathbf{x}_i] + b)$

where $\mathbf{x}_i$ are conventional features (momentum, value, quality) and $\|$ denotes concatenation.

Relational Risk Propagation

Supply-chain and credit linkages propagate risk across connected entities. An influence score from entity $j$ to $i$ through path $p$ is:

$\text{Influence}(j \to i) = \sum_{p \in \text{Paths}(j,i)} \prod_{(u,r,v) \in p} w_{u,r,v}$

where $w_{u,r,v}$ are edge weights reflecting the strength of each relationship.

Event Propagation on KGs

Financial events (earnings surprises, regulatory actions, M&A announcements) propagate through the knowledge graph. The propagated impact on entity $i$ from an event at entity $j$ is:

$\text{Impact}_i = \sum_{j \in \text{EventSources}} \alpha_{ij} \cdot \text{EventScore}_j$

where $\alpha_{ij}$ is the attention-weighted influence derived from the graph structure.

Feature Engineering

Entity Centrality Features

• Degree centrality: Number of connections (higher degree = more systemic importance)
• Betweenness centrality: Fraction of shortest paths passing through an entity
• PageRank: Recursive importance measure based on incoming link quality
• Eigenvector centrality: Influence measure based on connection to other influential nodes

Relationship-Based Features

• Supply chain depth: Distance to end consumer or raw material source
• Sector concentration: Diversity of cross-sector connections
• Peer similarity: Cosine similarity of KG embeddings between companies
• Contagion score: Weighted sum of distressed neighbors’ risk metrics

Temporal Graph Features

• Edge formation rate: Speed at which new relationships appear
• Relationship stability: Duration of existing connections
• Graph density change: Temporal evolution of local clustering coefficient

Applications

Supply Chain Alpha

Knowledge graphs mapping supplier-customer relationships enable lead-lag trading strategies. When a major supplier reports strong earnings, connected downstream companies often experience delayed positive price reactions, creating exploitable alpha signals.

Contagion-Aware Risk Management

By propagating credit risk through the KG, portfolio managers can identify hidden exposures. A seemingly diversified portfolio may have concentrated supply-chain risk that only becomes visible through graph analysis.

Event-Driven Trading

KGs enable systematic propagation of event signals. An FDA approval for a pharmaceutical company propagates to its suppliers, competitors, and partner companies with quantifiable impact weights derived from the graph structure.

Rust Implementation

The Rust implementation provides five core components:

1. KnowledgeGraph: Triple store with entity and relation indexing, neighbor lookup, and path finding
2. TransEModel: TransE embedding model with margin-based ranking loss and SGD training
3. GraphFeatureExtractor: Computes centrality features (degree, PageRank) and peer similarity from the KG
4. TradingSignalGenerator: Combines KG embeddings with market features to generate buy/sell signals
5. BybitClient: Async market data client for Bybit V5 API (klines and orderbook)

Bybit API Integration

The implementation fetches live market data from Bybit:

• Kline endpoint: /v5/market/kline — OHLCV candlestick data for feature computation
• Orderbook endpoint: /v5/market/orderbook — real-time bid/ask levels for spread features

Both stock market symbols (via KG entity mapping) and crypto pairs (BTCUSDT, ETHUSDT) are supported.

References

1. Bordes, A., et al. “Translating Embeddings for Modeling Multi-relational Data.” NeurIPS, 2013.
2. Schlichtkrull, M., et al. “Modeling Relational Data with Graph Convolutional Networks.” ESWC, 2018.
3. Cheng, D., et al. “Knowledge Graph-Based Event Embedding Framework for Financial Quantitative Investments.” AAAI, 2020.
4. Feng, F., et al. “Temporal Relational Ranking for Stock Prediction.” ACM TOIS, 2019.
5. Ding, X., et al. “Knowledge-Driven Stock Trend Prediction and Explanation via Temporal Convolutional Network.” WWW, 2019.