SNSF Narrative Digital Finance

Abstract

Large fluctuations, instabilities, trends and uncertainty of financial markets constitute a substantial challenge for asset management companies, pension funds and regulators. Nowadays, most asset management companies and financial institutions follow a so-called systematic trading approach in their investment decisions. Systematic trading refers to applying predefined, rule-based trading strategies for buy- and sell orders. However, automated or rules-based trading activities bring certain risks for market participants and the whole financial market. In times of increased market volatility, market turmoil or so-called market sell-offs, investors applying similar trading rules might undertake the same actions, escalating and increasing systemic market risk through such behavior. Such situations have been frequently observed on financial markets for instance, in March 2020 (sell-off related to the Covid pandemic), during the European Sovereign Debt crisis and the global financial crisis 2007-08. Research in economics and management has begun to embrace the role that narratives play in guiding individual and collective decision-making. McCloskey (2011) describes unforeseen growth in economic development yet goes on to explain that no economic theory is able to capture this extent. She argues that a change in rhetoric had basically freed a social class (the bourgeoisie) and given it a sense of dignity and liberty. As such, economic change, she argues, depends to a great extent on social narratives that shape ideas and the beliefs of people. Yet, despite the notion that narratives, individual and collective actions, and market outcomes are inextricably linked, our knowledge about the mechanisms or processes through which they interact and how narratives can inform opinions or sway current thinking is still evolving. Entrepreneurs, for example, may use verbal communication to achieve plausibility (i.e., generate the sense that a given interpretation of events appears acceptable) or resonance (i.e., obtain alignment with the beliefs of the target audience; see van Werven et al., 2019). They may do so through rhetoric such as storytelling (Navis & Glynn, 2011) or crafting compelling arguments (van Werven et al., 2015) as well as employing combinations of figurative language and gesturing (Clarke et al., 2021) as they manage and conform with the expectation of their audience.

Outcomes of invoking narratives are consequential. The literature has indeed documented various forms of verbal communication–including written texts such as social media posts and blogs, or business plans or spoken text (Garud et al., 2014; Clarke et al., 2019, Clarke et al., 2021) – as a crucial means to secure support and investment. The narratives or rhetoric employed in these stories are used as vehicles for assembling and communicating details about ideas and future possibilities (Garud et al., 2014). In summary, narratives help audiences make sense of situations and situate the description into the audience's social and cultural framework (Lounsbury and Glynn, 2001).

In the following, we, therefore, explore computational techniques to predict financial market outcomes using text, speech, and video/picture data. Advances in data processing and machine learning allow new ways of analysing data and may have profound implications for empirical testing of lightly studied, yet complex, empirical financial relationships. This project therefore integrates various forms of narratives into the context of financial market analysis, leverages machine learning techniques, and aims to show how narratives are inextricably interwoven in the continuously unfolding financial market evolutions. We will extend quantitative research through novel measurement techniques, the creation of new data sets, offering new solutions towards prediction problems, and the induction of new theories (Obschonka & Audretsch, 2020). We will also contribute to recent works that demonstrated the potential of theoretical and methodological advancements through the application of machine learning in the research practice (Mullainathan & Spiess, 2017; von Krogh, 2018). In pursuit of both practical 'relevance' of our research (Wiklund et al., 2019) and the contribution of "AI-integrated" research (Levesque et al. 2020), our approach will provide actionable insights.

Grant Link

Approach

In a first step, we will design a tool allowing us to collect all relevant data from various data sources. Indeed, collecting purely financial data, such as stock prices or macroeconomic indicators, can be easily performed using subscription-based platforms such as Bloomberg, Reuters or Investing.com. However, textual data will constitute a substantial challenge in terms of (i) collecting from the web, (ii) formatting, and (iii) pre-processing, including dating and categorising. For this purpose, we will develop an automated tool which will collect textual data, categorise them, date and store them in an easy to analyse format. We will manage our database with SQL solutions. The second step will focus on our research questions and the four building blocks listed below.

We will formulate numerous data-driven general/main and block-specific research questions within our hypothesis-driven project. The main research questions will be:

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

Block 1: Text data & text analytics

Text mining techniques are frequently used in scientific research for forecasting developments of various financial assets such as FX, equities, bonds, commodities see, for instance, Fung et al. 2003, Hajizadeh et al. 2010, Nassirtoussi et al. 2014, Kumar and Ravi 2016, Loughran and McDonald 2016 (Chan and Franklin 2011, Cambria and White 2014, Xing et al. 2017, Chen et al. 2020). Our solution uses NLP and text mining techniques for asset allocation and prediction to apply for structural breaks and change point detection combined with asset allocation methodology. For instance, those techniques are used for predicting cryptocurrency price bubbles using social media data (Biessey 2021). However, the field of classic financial assets tends to be under-researched. Therefore, we will collect relevant literature, review solution and answer the following research questions:

1. How can textual analysis and the application of natural language processing techniques be efficiently used for portfolio management, including risk management and asset allocation?
2. What are the most promising NLP / text analysis techniques?

Block 2: Structural breaks detection & asset price bubbles

The survey of econometric tests of asset price bubbles shows that, despite recent advances, econometric detection of asset price bubbles cannot be achieved with a satisfactory degree of certainty (Gürkaynak 2008). Furthermore, currently, there exist a relatively low number of scientific papers about the live detection of structural breaks in a systematic way. Most of the existing solutions have not been validated on real-world data. An obvious downside of such experiments is that the dynamics of the simulated data are often particular to the paper, and any model that corresponds to these dynamics has an unfair advantage (van den Burg and Williams 2020). Hence, we will tackle the issue of structural breaks and asset price bubbles in the three steps. In the first step, we will focus on post- ante structural detection methods for asset price bubbles to identify past breaks in real macroeconomic and financial time series. The breaks will be compared and based on consensus simplified if needed. In the second step, we will reapply well-established, known methods for live detection of breaks and check their ex-ante performance. Based on the current state of literature, we expect a relatively poor forecastability of breakpoints. Therefore, in the third step, we will involve NLP and text analysis techniques as a supporting or main method for detecting breakpoints. Within our research in this block, we will answer the following research questions:

1. How to detect, identify and date structural breaks in online and offline matters?
2. Detection of structural breaks / change points / asset prices bubbles in a live-matter using most recent (alternative) data (Twitter, News etc.)

Block 3: Narratives for structural breaks

The so-called "narratives block" will be highly dependent on results from block 1 and 2. Using newly acquired knowledge and experience with text analysis and NLP and insights into detecting structural changes, we will develop a framework with market narratives for detecting asset price bubbles. Narratives "go viral" and spread worldwide with economic impact (Shiller 2017). There is considerable evidence in the scientific literature showing that people respond strongly to narratives in the fields of marketing (Escalas 2007); journalism (Machill et al. 2007 ); education (McQuiggan et al. 2008); health interventions (Slater et al 2003); and philanthropy (Weber et al. 2006). We will answer the following research questions:

1. Can market narratives help predict financial market bubbles and their bust?
2. Can market narratives help detect financial market bubbles?
3. Can narratives sway investment opinions?

In order to achieve this, one needs to devise experiments for measuring narratives and then incorporate those measurements into predictive models aimed at explaining different aspects of financial market behaviour. Looking at the 3rd research question within this block, we can investigate the effectiveness of specific components of narrative strategies by carrying out an experiment with potential investors who will be given different investment options that: a) incorporate a narrative structure and b) emphasise different positive or negative emotions in the text. Put differently, the participants will have text and information presented to them, where the level of the manipulated component (e.g. emotional content) is low/high and a narrative appears/does not appear in the description of the investment. Once the data is collected, we can employ pre-trained Transformer models (GloVe Embeddings with Long Short-Term Memory Network (GVEL) or Bidirectional Encoder Representations from Transformers (BERT)) and finetune them to model the presence of narratives. This will also enable us to test how text impacts investment option perceptions by using text matching approaches that employ lower dimension summaries of texts (Roberts, Stewart, and Nielsen, 2020) or use low-dimensional representations as causally sufficient embeddings (Veitch, Sridhar, and Blei, 2020). Such experiments will provide evidence of the causal influence that different components of narratives and emotions exert on the appeal of these investments by influencing the judgements of potential investors. Furthermore, for narratives to be effective they need to resonate with potential investors (van Werven et al., 2019). We thus use a study to test the effectiveness of specific components of narrative strategies in an experimental design. We will introduce participants to different investment options that a) incorporate a narrative structure b) emphasise different positive or negative emotions in the text. Due to practical considerations (sample size), we will employ a 2x2 experimental design for each of the components tested (e.g. narrative element present vs. purely informational text; emotionality of communication low vs. high). The experiments will provide evidence of the causal influence that different components of narratives and emotions exert on the appeal of these investments by influencing the judgements of potential investors. This experiment might also inform us about narratives and biases, where for example, investors forego certain investment options (overlook correlations that may improve their risk-return profile in the portfolio) because of the narratives presented to them.

Block 4: Multidimensional AI and ML solutions in a fully integrated framework

As already mentioned in the literature review, AI and ML techniques possess a substantial potential to revolutionise financial markets (Milana and Ashta 2021). New technologies transform business models and markets for trading, credit and blockchain-based Finance, generate efficiencies, reduce friction, enhance product offerings, and refine the existing financial services industry (Buchanan 2019, Hilpisch 2020, Moloi and Marwala 2020). Since, in previous blocks, we look at detecting structural breaks and asset price bubbles from various perspectives and apply different techniques, it seems to be self-explanatory and expected to check if those methods can be combined into a fully integrated framework. Research questions:

1. Can a combined ML approach outperform each single method?
2. Do complex AI and ML approaches outperform simple forecast combinations?

Relevance and impact

Scientific relevance

Expected impacts for research in Finance

Potentially, the impact on Finance will be substantial. First and foremost, our approach will help to better understand financial markets, the role and impact of different market participants and how the dynamics of markets arise and can be explained. Our research can lead to a significant transformation of the way we model many of the research questions in Finance, most notably any that are related to risk management, new techniques and modelling, particularly in all the major fields of quantitative Finance, such as modelling financial markets, risk management, systemic risks, asset pricing, portfolio optimization, trading strategies, but also

political considerations and the interplay of market participants that have a substantial impact on markets such as governments and central banks. As a prime example, we expect that our academic results can help the ECB and governments to explain the dynamics of financial markets, and certain irrationalities and mitigate future risks. Furthermore, traditional linear econometric models are more and more challenged with our ever-more complex, non-linear world and our approach will help to augment classical theories and give insights from a different angle, reducing the dependency on models and their associated risks, that have shown up several times in the previous years, such as the financial crisis (2007-2009) or the European debt crisis in Europe (2010-2011).

Broader scientific impact

Impact for research in other related domains.

Our results will show how to build a structured, consistent framework for structural breaks detection using various methods and various data sets. On top of numerous approaches, we will propose a machine learning or artificial intelligence-based technique which would help to combine all information in a consistent manner. Since our entire approach is heavily driven by data and quantitative research questions, our research can impact other fields than Finance. We will research and investigate various approaches for text analysis which can be useful for any field requiring working with textual data. The detection of structural breaks or of any highly negative events can be easily transformed to various political and social science research activities.

Impact for education and teaching.

As it pertains to the development of PhD students, the hired PhDs working on this project will be integrated into an international, interdisciplinary and intersectoral network community of experts providing training and collaboration opportunities for a broad range of skills. For education and teaching, the research will be integrated into PhD courses on data science, machine learning and quantitative Finance. Specifically, we are planning to have a first PhD course available by spring 2023, that combines all the interdisciplinary fields we are considering here. Also, we are planning to organize several training schools for PhD students and young researchers, investigating the interdisciplinary fields of our research. The first one will take place in Switzerland in September 2023 and will be jointly with several participants of our COST Action, most notably Profs. Poti, Härdle, Hirsa, Marazzina. The research results will also be integrated into a new European Industrial Doctorate on financial data science, that we are planning jointly with University College Dublin, Humboldt-University Berlin, University of Bucharest, to be submitted to the Marie Sklodowska- Curie Action in the European Horizon Europe framework programme. We are confident that this project might lay the foundation not subsequently help to inspire a next generation of finance researcher. Early career researchers will find themselves in the position where they will develop theoretical and technical skills that allow them to contribute expertise at the intersection of finance and data science, but that also will enable them to support the research of other network partners and scientifically contribute to theories and methods to neighbouring areas and research fields.

Publication of research results.

We aim to publish our results in some of the following journals: The Journal of Portfolio Management, The Journal of Financial Data Science, European Journal of Operational Research, International Journal of Investment Management and Financial Innovations, Journal of Mathematical Finance, The Journal of Finance, Review of Financial Studies, Journal of Financial Economics, Encyclopedia of Financial Models. The results will be disseminated at large international conferences. We have already accepted an invitation for the 8th European COST Conference on Artificial Intelligence in Finance and Industry in Switzerland in 2022 and the SIAM Annual Meeting in 2022. We will also write white papers for industry and policy papers for regulators.

Broader impact

Need of research for finance practitioners/industry.

The finance industry is challenged by high risks, ongoing and repetitive financial crises and market efficiencies. Despite many years of financial markets research, substantial regulatory investments and efforts over the last 20 years, we are far from a sufficient understanding of financial markets. Our research will help to improve the general understanding of financial markets. The research question is directly defined and shaped by the Finance industry (JP Morgan, Credit Suisse, UBS, Goldman Sachs), looking into new academic models on MARL to improve their trading strategies and risk management policies. As for reinforcement learning specifically, it has transformed several practice areas, but not yet in Finance, despite various ongoing efforts. We aim to solve the practical problem of optimal risk management during financial shocks in global markets by producing scientific insights concerning the behaviour of differently motivated agents in financial ecosystems.

Ensuring implementation of research results

Structural break detection, their validation and precise dating is a topic of significant importance for the financial industry.. Our results can be directly put into practice by testing with new datasets that banks have access to and deploying the simulation engine to the IT environment of banks. Banks can, e.g. simulate the impact of their trading behaviour, conditioned on the actions of other significant players. Regulators, central banks and governments can apply the methodology to simulate stress scenarios, imbalances in the market and systemic risks. We have extensive industry contacts, such as with the research teams from ING Group, Credit Suisse, UBS, and multi-national organizations such as the ECB and ESM. All of them are explicitly interested in this research and have expressed their wish to be involved. A near-term implementation of the results and a continuation via an Innosuisse project (Swiss-based use-inspired funding) is not only expected but already firmly planned. About 50% of our research projects are use-inspired, and we have vast experience implementing our results within the industry. Changes outside science and their nature. Our research also has the potential to be particularly relevant for governments, central banks and regulatory authorities. We will work with BIS and about 10 European regulators (from our COST and EU H2020 network) to show the results. In addition, we will write two policy papers and two position papers on a) how to better model asset price bubbles in financial markets and b) steps and measures to undertake that can lead to a more resilient financial system and fewer risks. From a citizen's perspective, we expect our results to lead to a) a more stable financial system with fewer imbalances and risks and b) improved investment products with better performance and a more efficient finance industry. In addition, the European Commission has recently published three strategy plans on the capital market union, the approach to artificial intelligence, and a digital strategy. All three show the need for more quantitative solutions and more focus on artificial intelligence methods.

In line with COST Goals.

Our research setting will also substantially contribute to additional goals of the COST Action 19130. We have envisioned many measures of knowledge transfer and cooperation and the training and support of young researchers. We will use the COST network to disseminate our results and have an impact on policymakers (e.g. by exchanging ideas with DGFISMA: Directorate General Financial Markets of the European Commission) We organise an annual conference and a monthly research seminar in Switzerland. In addition, we use the short-term scientific missions to invite researchers to Switzerland and visit researchers and companies abroad, all within the context of Action 19130. As a flagship event, we will host the annual management committee meeting at our university in September 2023. We expect more than 100 researchers from all over Europe to join us for one week of research events.

References

1. Cambria, Erik, and Bebo White. "Jumping NLP Curves: A Review of Natural Language Processing Research." IEEE Computational Intelligence Magazine 9, no. 2 (2014): 48–57.
2. Chan, Samuel WK, and James Franklin. "A Text-Based Decision Support System for Financial Sequence Prediction." Decision Support Systems 52, no. 1 (2011): 189–98.
3. Chen, Chung-Chi, Hen-Hsen Huang, and Hsin-Hsi Chen. "NLP in FinTech Applications: Past, Present and Future." ArXiv Preprint ArXiv:2005.01320, 2020.
4. Escalas, Jennifer Edson. "Self-Referencing and Persuasion: Narrative Transportation versus Analytical Elaboration." Journal of Consumer Research 33, no. 4 (2007): 421–29.
5. Fung, G. Pui Cheong, J. Xu Yu, and Wai Lam. "Stock Prediction: Integrating Text Mining Approach Using Real-Time News." In 2003 IEEE International Conference on Computational Intelligence for Financial Engineering, 2003. Proceedings., 395–402. IEEE, 2003.
6. Garud, R., Schildt, H. A., & Lant, T. K. (2014). Entrepreneurial storytelling, future expectations, and the paradox of legitimacy. Organization Science, 25(5), 1479-1492.
7. Gürkaynak, Refet S. "Econometric Tests of Asset Price Bubbles: Taking Stock." Journal of Economic Surveys 22, no. 1 (2008): 166–86.
8. Kumar, B. Shravan, and Vadlamani Ravi. "A Survey of the Applications of Text Mining in Financial Domain." Knowledge-Based Systems 114 (2016): 128–47.
9. Loughran, Tim, and Bill McDonald. "Textual Analysis in Accounting and Finance: A Survey." Journal of Accounting Research 54, no. 4 (2016): 1187–1230.
10. Lounsbury, M., & Glynn, M. A. (2001). Cultural entrepreneurship: Stories, legitimacy, and the acquisition of resources. Strategic management journal, 22(6‐7), 545-564.
11. Machill, Marcel, Sebastian Köhler, and Markus Waldhauser. "The Use of Narrative Structures in Television News: An Experiment in Innovative Forms of Journalistic Presentation." European Journal of Communication 22, no. 2 (2007): 185–205.
12. McCloskey, D. N. (2011). The Rhetoric of the Economy and the Polity. Annual Review of Political Science, 14, 181-199.
13. McQuiggan, Scott W., Jonathan P. Rowe, Sunyoung Lee, and James C. Lester. "Story-Based Learning: The Impact of Narrative on Learning Experiences and Outcomes." In International Conference on Intelligent Tutoring Systems, 530–39. Springer, 2008.
14. Nassirtoussi, Arman Khadjeh, Saeed Aghabozorgi, Teh Ying Wah, and David Chek Ling Ngo. "Text Mining for Market Prediction: A Systematic Review." Expert Systems with Applications 41, no. 16 (2014): 7653–70.
15. Roberts, M. E., Stewart, B. M., & Nielsen, R. A. (2020). Adjusting for confounding with text matching. American Journal of Political Science, 64(4), 887-903.
16. Shiller, Robert J. "Narrative Economics." American Economic Review 107, no. 4 (2017): 967–1004.
17. Slater, Michael D., David B. Buller, Emily Waters, Margarita Archibeque, and Michelle LeBlanc. "A Test of Conversational and Testimonial Messages versus Didactic Presentations of Nutrition Information." Journal of Nutrition Education and Behavior 35, no. 5 (2003): 255–59.
18. Van Werven, R., Bouwmeester, O., & Cornelissen, J. P. (2015). The power of arguments: How entrepreneurs convince stakeholders of the legitimate distinctiveness of their ventures. Journal of Business Venturing, 30(4), 616-631.
19. Veitch, V., Sridhar, D., & Blei, D. (2020, August). Adapting text embeddings for causal inference. In Conference on Uncertainty in Artificial Intelligence (pp. 919-928). PMLR.
20. Weber, Keith, Matthew M. Martin, Members of COMM 401, and Michael Corrigan. "Creating Persuasive Messages Advocating Organ Donation." Communication Quarterly 54, no. 1 (2006): 67–87.