Диссертации, представленные на защиту и подготовленные в НИУ ВШЭ

Прогнозирование финансовых потоков является ключевой задачей для банковского сектора, инвестиционных компаний и регуляторов. Современные методы машинного обучения, в частности глобальные модели, демонстрируют высокую прогностическую точность, однако их применение ограничено чувствительностью к качеству данных, недостаточной разработанностью универсальных подходов к интеграции разнородной неструктурированной информации (в том числе финансовых новостей) и низкой интерпретируемостью прогнозов. В работе предложены эффективные процедуры предобработки временных рядов, способы генерации текстовых признаков, позволяющие выделять значимые сигналы из новостного фона и учитывать их влияние на динамику финансовых потоков, подходы к построению глобальных моделей, а также подходы к разработке мульти-модальных моделей, объединяющих временные ряды и текстовые данные при сохранении высокой интерпретируемости. Эффективность предложенных методов подтверждена на масштабных данных российского финансового сектора. Полученные результаты имеют существенную практическую значимость для оптимизации налично-денежного обращения, управления финансовой ликвидностью и разработки инвестиционных стратегий.

Monetary instability is a persistent concern for emerging market economies, especially due to their vulnerability to external financial shocks. These shocks often reveal structural weaknesses in how these economies are positioned within the global financial system. This dissertation examines two important sources of such instability: changes in U.S. interest rates and uncertainty in U.S. economic policy. It first investigates the effects of these external factors on macroeconomic outcomes in emerging markets. The analysis then turns to the domestic level, using India as a case study to examine how money demand can be forecast more accurately. A panel structural VAR is used to study the effects of U.S. interest rate shocks, while a GMM-based panel VAR examines the impact of U.S. economic policy uncertainty. The results indicate that both higher interest rates and greater policy uncertainty in the U.S. are associated with exchange rate depreciation, rising inflation, and tighter financial conditions in emerging markets. In the case of India, the dissertation compares several machine learning models for forecasting money demand and finds that methods like LSTM and LASSO perform particularly well. These findings highlight distinct sources of monetary instability and offer practical insights that may be useful for both researchers and policymakers.

This thesis explores the intersection of human-AI interaction and transportation forecasting, focusing on the development and application of attention-based neural network models, particularly memory networks. The research investigates the challenges inherent in traffic prediction tasks, such as capturing complex spatio-temporal dependencies and providing interpretable predictions. Leveraging the attention mechanism, the proposed models aim to enhance prediction accuracy and provide insights into the underlying factors influencing traffic dynamics. The thesis encompasses a comprehensive analysis of spatio-temporal data, model development, and evaluation, as well as techniques for visualizing and interpreting model predictions. Through empirical studies and case examples, the effectiveness and practical implications of the proposed approach are demonstrated. The findings contribute to advancing the understanding of human-AI interaction in transportation systems and provide valuable insights for improving the accuracy and interpretability of predictive models in complex real-world scenarios.

The quick development of the internet and information technology (IT) worldwide has given access to a large amount of data, which are usually known as “big data". One of the main tools that can be used to analyze big data is a search engine, which is often considered the first step in the consumer decision-making process, to understand social dynamics and to make better predictions. In this regard, Google is the search engine with the largest market share worldwide (90% in 2018) and the analysis of its search data has been one the most important and well-known examples of the use of big data. In 2006, it launched a tool named Google Trends which shows how frequently a particular keyword or a topic are searched online in a specific region, at a specific period of time, and also in different languages. The advent of these large datasets further stimulated the interest in developing multivariate models able to consider departures from the assumption of normality and to be computationally tractable: Copula models can deal with both these two issues. The theory of copulas dates back to Hoeffding (1940) and Sklar (1959), but its large-scale use in empirical applications is far more recent and dates back to the first decade of the 21st century. Copulas allow for a flexible modelling of the dependence structure between different variables, as well as for the possibility to have different marginal distributions. Moreover, the separation of the dependence structure from the marginals strongly decreases the computational burden of estimating a multivariate model. The 20 publications that constitute my dissertation investigated several cases where it is possible to model and forecast economic and financial time series using Google data and/or copula models.