FinDKG: Financial Dynamic Knowledge Graph

This website provides the Financial Dynamic Knowledge Graph (FinDKG) portal, driven by graph AI model KGTransformer, from the streams of global financial news. Should you use the data or model, kindly cite the paper.

FinDKG Snapshot as of Week @Jan 1, 2023

The interactive graph below presents the FinDKG's key entities with their actively linked entities - the most influential variables driving the latest global macro-financial universe (drag the interested entity to see impact):

The follwing FinDKG analytical table provides a detailed profile of these "Top KG Entities", alongside financial variables expected to be most influenced by them in the coming week, as predicted by our graph-based AI model:

The column labeled "Predicted Most Impacted Financial Entities" identifies the financial variables that are forecasted to experience the most significant changes in the week ahead. Additionally, the table ranks these entities within a vast universe of over 10,000 other entities, using the “Rank Percentile” score (0% to 100%) to signify their relative importance of the FinDKG. A separate "Novelty" z-score is included to highlight how recently the entity has appeared in the dataset. Moreover, by scrutinizing the "Recent 3-month Trend" column, users can discern the evolving influence trend of a particular entity over the past quarter.

Open-Source Financial Dynamic Knowledge Graph Dataset

The financial sector, despite its wealth of text data, often lags in the availability of specialized knowledge graphs (KGs). Addressing this need, we introduce the FinDKG dataset — a unique, open-source Financial Dynamic Knowledge Graph, designed to provide a temporally-resolved representation of global economic and market trends. Leveraging curated open-source financial news articles of deep history within Wayback Machine archive, FinDKG serves as an invaluable resource for researchers. The dataset is research available (non-commercial purpose only) for download on GitHub.

Bridging the Gap in Finance-Specific Knowledge Graphs

Knowledge graphs have proliferated in various domains, but their presence is notably sparse in sector-specific areas like finance. The FinDKG dataset aims to bridge this gap by providing a rich, temporally-resolved knowledge graph specifically crafted for the financial domain.

FinDKG Schema and Relationships

FinDKG features a well-defined schema that includes 15 predefined relationships and 12 predefined entity types. From modeling the causal effects between financial markets to tracking influential persons and impactful events, FinDKG provides a comprehensive framework for understanding financial dynamics.

Use FinDKG Dataset for Your Research

The FinDKG dataset is organized to facilitate easy usage, including train/validation/test splits organized chronologically. It suits perfectly for extrapolative Temporal Link Prediction task. To get your hands on this invaluable resource, simply head over to our GitHub repository for the dataset download.

KGTransformer for Dynamic Knowledge Graph Learning

The rise of modern Knowledge Graphs (KGs) in 2012 was a milestone that shifted the focus from keyword-based search to a more nuanced, context-aware understanding. However, the majority of the research and applications still rely on static KGs, overlooking their dynamic nature. To address this gap, we introduce the KGTransformer, a cutting-edge Graph Neural Network model tailored for Temporal Knowledge Graphs Learning. KGTransformer not only excels in graph analytics when compared to static KG models, but it also offers a powerful framework for fine-grained temporal inference and dynamic pattern recognition. Interested readers can delve into the technical architecture and get access to the KGTransformer source code via GitHub repository.

Model Architecture

The KGTransformer extends the existing Graph Attention Network (GAT) to incorporate the KG "meta-relations." In simple terms, it enables the model to focus not just on the relationship but also on the type of entities involved. This enhanced focus is achieved through an extended attention mechanism that brings an unprecedented level of nuance and contextualization to node representations within the graph.
     For instance, if OpenAI invents a new model like ChatGPT, KGTransformer could capture various aspects of this relationship—from financial transactions to technological synergies—all simultaneously.

Temporal Knowledge Graph Learning

Temporal Knowledge Graphs (TKGs) add a time dimension to traditional static knowledge graphs, capturing the evolving relationships between entities across various domains. This added complexity presents new challenges in graph analytics, such as temporal link prediction, that static models simply aren't equipped to handle. TKGs associate a time-stamp with each edge in the graph, providing a rich temporal context for understanding how relationships between entities change over time.
     Unlike traditional models, KGTransformer features a probabilistic dynamic graph learning framework that allows it to adapt to the time-evolving nature of TKGs. It utilizes specialized attention mechanisms that are capable of recognizing and learning both the structural intricacies and the temporal dynamics inherent in these complex graphs. By leveraging both entity and relationship metadata, KGTransformer provides a more nuanced and accurate analysis for tasks like temporal link prediction, setting a new standard for how we understand and interact with evolving knowledge structures.

Performance Metrics & Real-world Application

Out-of-sample evaluation results on the FinDKG dataset show a remarkable performance uplift of approximately 15% in key TKG prediction metrics like Mean Reciprocal Rank (MRR) and Hits@3,10 when compared to other state-of-the-art TKG models. These metrics validate KGTransformer's superiority in capturing the dynamic interrelations within TKGs.

ICKG: Integrated Contextual Knowledge Graph Generator Large Langue Model

Integrated Contextual Knowledge Graph Generator, termed ICKG, is an innovative, instruction-following Large Language Model (LLM) that is fine-tuned for the purpose of generative knowledge graph construction. Drawing from the advanced capabilities of LLM and built on top of the Vicuna-7B from Meta's LLaMA 2.0 models, ICKG offers unparalleled proficiency in converting vast amounts of unstructured textual data into structured knowledge graphs. With proven efficiency for deployment on consumer-level GPUs, ICKG not only matches the performance of leading models but also does so at a fraction of the computational cost. This makes ICKG an essential tool for researchers and professionals aiming to unlock actionable insights from knowledge graph data across various domains.

To train ICKG, we began with Vicuna-7B as our foundational architecture. Building on this, we fine-tuned ICKG using self-instruct methodology involving GPT-4's API to generate high-quality instruction-response pairs. The process involved 3K demonstrations that were specifically engineered to guide the model in knowledge graph construction. This fine-tuning allows ICKG to excel at extracting entities and relationships from unstructured text data, transforming it into structured knowledge graphs efficiently and accurately.

The Challenge of Knowledge Graph Construction: Converting Text into Structured Data

In an age where text data is growing exponentially, the biggest challenge isn't just collecting it, but structuring and making sense of it. Knowledge graphs, which can represent complex relationships between entities, have become a cornerstone in tackling this challenge. They've found applications in a wide array of sectors, from search engines to healthcare and finance. However, constructing these graphs manually is time-consuming, and automating this process requires sophisticated technology.

Leveraging LLMs with Prompt Engineering

The beauty of ICKG lies in its reliance on Large Language Models (LLMs) and a technique called "prompt engineering." LLMs like GPT-4 have the capability to understand and generate human-like text, thanks to their training on extensive text data. ICKG is fine-tuned from Vicuna-7B, which itself is based on Meta's LLaMA model, and utilizes the GPT-4 API for data generation.

Prompt engineering is where the magic happens. It involves manipulating the input query or "prompt" to guide the model's response effectively. With this method, ICKG can precisely extract intricate relationships and entities from text, making it incredibly effective for knowledge graph construction.

FinDKG System: Graph AI for Global Financial Systems

FinDKG system is a revolutionary open-source graph-based AI system designed to model the global financial systems through dynamic Knowledge Graphs (KGs). Powered by the KGTransformer, a state-of-the-art Graph Neural Network model, FinDKG provides forward-thinking solutions for broad financial applications like risk management, thematic investing, and economics forecasting. Unlike traditional financial models and static KGs, FinDKG offers the advantage of temporal adaptability, dynamic expressiveness, and intelligent reasoning. In essence, FinDKG leverage KG as the alternative datasets and cutting-edge graph learning models, goes beyond conventional financial analytics to anticipate market trends, manage risks, and identify investment opportunities in a way that has never been done before.

Risk Management with FinDKG - A Covid-19 Risk Tracking Case Study

The Covid-19 pandemic presented an unprecedented challenge for financial risk management, affecting economies and markets globally. Using FinDKG's dynamic capabilities, we effectively modeled the pandemic's ripple effects throughout the global financial systems. By employing rolling 1-month snapshot knowledge graphs updated weekly, FinDKG was adept at capturing the pandemic's evolving nature. The graph's composite centrality measures highlighted significant periods such as the initial outbreak and vaccine releases, providing invaluable insights for financial decision-makers. FinDKG's robust analytical framework thus proves invaluable for tracking systematic risks in complex contagion scenarios like Covid-19.

FinDKG-powered Thematic Investing: Quantifying Forward-looking Company AI Exposures

Thematic investing targets macro-level trends like artificial intelligence or clean energy. Traditional methods often lack dynamic, forward-looking data. FinDKG, empowered by KGTransformer, addresses this by offering real-time, predictive insights into companies genuinely exposed to these themes.
     We showcase how FinDKG can deliver insights in AI-themed investing by capturing both supply and demand dynamics. It identifies AI-creating companies by translating future relation prediction as temporal link prediction, represented as (?, produce, AI), and those impacted by AI, tagged as (AI, impact, ?). This dual lens offers investors a comprehensive view of AI's market influence, enabling more informed investment decisions.

A Forward-looking Edge via Dynamic Knowledge Graph Lens

In the sphere of financial analytics, FinDKG represents a significant paradigm shift. Whereas traditional models are often retrospective, tethered to historical data, FinDKG leverages the temporal learning capabilities of KGTransformer to offer proactive, forward-looking intelligence. This enables FinDKG to adapt seamlessly to the fluid dynamics of financial markets and anticipate disruptions or opportunities with a level of foresight rarely seen in conventional models. Such predictive acumen affords investors and strategists an invaluable tool for timely risk mitigation and strategic allocation, elevating FinDKG as an indispensable asset in navigating today's complex financial landscape.

FinDKG Frequently Asked Questions (FAQ)

General

Q: What is the FinDKG System?
A: FinDKG is an open-source system that leverages dynamic knowledge graphs for financial analytics. It is powered by the KGTransformer and aims to model the global financial system.

Q: What is the KGTransformer?
A: KGTransformer is the underlying fitted model powering FinDKG. It is a graph machine learning model designed to model temporal dependencies and dynamic interrelationships among financial entities.

Q: Who is FinDKG for?
A: FinDKG targets academic researchers in machine learning, statistics, and finance. Available under a non-commercial license, it's built for those interested in interdisciplinary, data-driven research. The author noted its potential in practical applications, he will develop industry-specific projects entirely separated.

Data

Q: What kind of data does FinDKG use?
A: FinDKG mainly utilizes global financial news to form its dynamic knowledge graphs.

Q: How often is the data updated?
A: The system is actively planned to continuously update and process future incoming data streams. Snapshot knowledge graphs are assembled every week on Sundays.