Research Objectives

Work Program & Methodology | SNSF Grant IZCOZ0_213370

Overview

The proposed work program comprises four work packages (WP) with a conceptually new approach compared to existing research in financial market analysis:

Our empirical setup studies the interplay of narratives, language evolution, financial innovation, and market performance in a comprehensive cross-layer framework where evolutions, causality, and interactions can be measured explicitly.

Main Research Questions
In what sense are financial markets (ex-ante) predictable?
Is the ex-ante forecastability persistent, can it be applied for real use cases and to which extent?
How can structural break detection and changes in financial time series improve and complement modern portfolio theory?
Work Packages

WP1: Text Data & Text Analytics

Text mining techniques are frequently used for forecasting developments of various financial assets (FX, equities, bonds, commodities). Our solution uses NLP and text mining techniques for asset allocation and prediction, combined with structural breaks and change point detection methodology.

While techniques exist for predicting cryptocurrency price bubbles using social media data, the field of classic financial assets tends to be under-researched.

Research Questions
How can textual analysis and NLP techniques be efficiently used for portfolio management, including risk management and asset allocation?
What are the most promising NLP / text analysis techniques?

WP2: Structural Breaks Detection & Asset Price Bubbles

Despite recent advances, econometric detection of asset price bubbles cannot be achieved with satisfactory certainty. Currently, there exists a relatively low number of scientific papers about live detection of structural breaks in a systematic way, and most existing solutions have not been validated on real-world data.

Three-Step Approach
  • Step 1: Post-ante structural detection methods to identify past breaks in real macroeconomic and financial time series
  • Step 2: Reapply well-established methods for live detection and check ex-ante performance
  • Step 3: Involve NLP and text analysis as supporting/main method for detecting breakpoints
Research Questions
How to detect, identify and date structural breaks in online and offline matters?
Detection of structural breaks / change points / asset price bubbles in live-matter using alternative data (Twitter, News etc.)

WP3: Narratives for Structural Breaks

Narratives "go viral" and spread worldwide with economic impact (Shiller 2017). There is considerable evidence that people respond strongly to narratives in marketing, journalism, education, health interventions, and philanthropy.

Methodology (Evolved)
  • Multimodal influence framework extending NLP to images, video, and audio for financial pricing
  • TOPol: Semi-unsupervised framework using transformer embeddings, UMAP, and Leiden clustering
  • BERTopic-based narrative shift analysis quantifying polarity drift across economic regimes
  • Semantic centroid movement and keyword evolution tracking at structural breakpoints

Note: Original 2x2 experimental design evolved toward the multimodal influence framework, providing a more comprehensive theoretical foundation.

Research Questions
Can market narratives help predict financial market bubbles and their bust?
Can market narratives help detect financial market bubbles?
Can narratives sway investment opinions?

WP4: Multidimensional AI and ML Solutions

AI and ML techniques possess substantial potential to revolutionize financial markets. New technologies transform business models and markets for trading, credit, and blockchain-based finance, generating efficiencies and refining financial services.

Since previous blocks examine structural breaks and asset price bubbles from various perspectives using different techniques, we check if these methods can be combined into a fully integrated framework.

Research Questions
Can a combined ML approach outperform each single method?
Do complex AI and ML approaches outperform simple forecast combinations?
Expected Outcomes

Unique Contributions

Our solution for detecting structural breakpoints will be unique: (1) Exclusive focus on ex-ante forecasting (live detection), easily adjustable to rapidly changing markets; (2) Combining existing data sources, developing new quantitative models, and new frameworks to understand markets.

Enhanced Prediction Models

Financial market models with greater accuracy in detecting asset price bubbles and structural breaks

Novel Datasets & Methods

Deeper insights into the interplay between market narratives and financial indicators

Practical Tools

For asset management and regulatory bodies to better anticipate and react to market crises

Data Sources

Our research utilizes diverse textual and financial datasets:

RavenPack

Financial news headlines for sentiment analysis and narrative detection

LSEG (Refinitiv)

Earnings call transcripts for corporate narrative analysis

BIS Gigando

Worldwide central bank speeches for monetary policy narratives

SEC EDGAR

10-K and 10-Q filings for regulatory text analysis

Deutsche Borse

Nanosecond-level Xetra/Eurex trading data for HFT research

St. Louis FED FRED

Macroeconomic indicators (CPI, GDP, Unemployment, FED Funds Rate)

Custom data pipelines developed for collecting, formatting, and pre-processing textual data. Infrastructure handles large-scale processing with dedicated computing resources.

Approach Summary

Step 1: Data Collection Tool

Custom data pipelines collect and structure datasets from RavenPack, LSEG, BIS, SEC EDGAR. Data categorized, dated, and stored for analysis. Market data from Deutsche Borse and FRED.

Step 2: Research Execution

Focus on research questions across four building blocks, formulating data-driven hypotheses and testing within each work package.