Various shocks and uncertainties raise new challenges of modelling in industry practice: advanced methods driven by academic research can potentially solve for these challenges. For instance, firm specific variables can behave very differently in various baseline and stress scenarios, hence, regime switching approaches may lead to smaller forecast errors compared to standard techniques. Besides that, there are also various other areas where new academic research can bring in new ideas into industry. For instance, statistical tests have to be adjusted due to small sample size, where the critical values are taken from academic papers. In this presentation a few examples and challenges are shown based on some findings from Morgan Stanley Finance. The regulatory modelling team is exploring various methods based on econometric and machine learning techniques to improve the forecasting power of models, and on the top of that the team works closely with universities. Hence, there are several approaches supporting synergies across industry and academic practice within the team.

As AI-generated summaries become increasingly integral to information dissemination, establishing their quality with precision is critical. This presentation goes beyond traditional validation metrics like ROUGE and BLEU, exploring more sophisticated techniques for a holistic evaluation of summary quality on Finance. We will explore industry-standard methods including semantic similarity measures and their limitations, introducing alternative approaches such as information density analysis, and question-answering frameworks for comprehensive validation. Drawing on empirical research and case studies, we aim to present a nuanced perspective on LLM summary evaluation, highlighting innovative methodologies that offer deeper insights into the summarization capabilities of AI models.

Several asset classes show mean-reverting features, e.g. commodities, commodity futures, long-term safe assets (gold). We investigate the portfolio choice problem for investors with exponential utilities (=high risk aversion) as the investment horizon T tends to infinity. It turns out that the optimal equivalent safe rate grows in a superlinear way, depending on the strength of the mean-reversion effect. We cannot find the exact optimisers but construct a family of simple, explicit strategies that are optimal asymptotically (they generate equivalent safe rates of the optimal order). Interestingly, the presence or absence of a drift leads to entirely different conclusions, the nonzero drift case spectacularly outperforming the driftless one. Time permitting, we also review some related results on fractional Brownian motion.

The Kennedy model uses Gaussian random fields to model forward rates, providing a natural solution for handling negative interest rates. In our research, we provided probability 1 and maximum likelihood estimations for the parameters of the Kennedy field using Radon-Nikodym derivatives. Additionally, we present an efficient method for simulating the Kennedy field. We derive Black-Scholes-like pricing formulas for various financial products (caplets, floorlets, and swaps). Furthermore, we introduce a parameter estimation algorithm based on numerical extreme value analysis (specifically stochastic gradient descent), which we test on different financial products, first using simulated data and then assessing the Kennedy field's fit to real par swap rates. Finally, we present a neural network built for calibration, with ongoing work to improve its accuracy.

XVA and in particular CVA (Credit Valuation Adjustment) has an invaluable literature, with various methodologies suitable for the different needs of the user. However, in a real-life situation, when implementing a solution in an existing pricing library, one has to face a few additional challenges related to design, data, legal agreements and has to meet regulatory requirements as well.

Geopolitical events impact volatilities of most assets, sectors, and countries. To this end measuring geopolitical risks has been of key interest in the literature as seen by the development of the COVOL index of Engle and Campos-Martins (2023). COVOL uses volatility innovations in multivariate equity index portfolios to quantify global risks. This study looks at how geopolitical risks influence the financial sector. In particular, we use the Diebold-Yilmaz approach to analyse how global uncertainty influences financial stress in the UK and Hungary. We also use a growth-at-risk framework to better understand what part of the growth distribution is impacted by global uncertainty. Finally, using stochastic differential equations we predict the global financial risk index on one month horizon and use machine learning methods to update the model parameters. The purpose of the prediction is to measure the vulnerability of growth in timely manner and to generate stress-test scenarios that account for global uncertainty explicitly.

Adjustable-rate mortgages (ARMs) introduce a delicate balance between borrower risk and bank preference due to their potential for increased monthly payments resulting from rising interest rates. While banks favor ARMs for their low duration, borrowers face uncertainty. Currently, banking practices assume a constant interest rate throughout the mortgage term, neglecting the dynamic nature of interest rate fluctuations. This paper advocates for a novel approach: factoring in the expected interest rate trajectory when computing monthly ARM payments. Our proposed theoretical model yields a formula that fosters smoother payment profiles, enhancing financial stability. This innovation simultaneously curbs default risk associated with ARMs, preserving their desirable low duration. Through this endeavor, we contribute to a refined understanding of ARM dynamics, reconciling bank advantage with borrower security.