Research

Publications

Portfolio optimization and risk management through Hierarchical Risk Parity and Logic Learning Machine: a case study applied to the Turkish stock market

RISK MANAGEMENT MAGAZINE Vol. 19, Issue 1 January – April  2024  (with Pier Giuseppe Giribone, Marco Muselli, Erenay Ünal, Damiano Verda)

Abstract

This study explores an innovative approach to portfolio optimization, bridging traditional Modern Portfolio Theory (MPT) with advanced Machine Learning techniques. We start by recognizing the significance of Markowitz's model in MPT and quickly proceed to focus on the Hierarchical Risk Parity (HRP) method. HRP overcomes some of the limitations of Markowitz's model, particularly in managing complex asset correlations, by offering a more refined risk management strategy that ensures balanced risk distribution across the portfolio. The paper then introduces an innovative Machine Learning approach that employs the Logic Learning Machine (LLM) method to enhance the explainability of the Hierarchical Risk Parity strategy. Such integration is considered the core research part of the study, given that its application makes the output of the model more accessible and transparent. A case study based on the Turkish stock market has been provided as an example. The combination of traditional financial theories with modern Machine Learning tools marks a significant advancement in investment management and portfolio optimization, emphasizing the importance of clarity and ease of understanding in complex financial portfolio models.

Working Papers

2025

Improving the Performance of Traditional Interest Rates Term Structure Models Using an Artificial Intelligence-Based Approach. Evidence from major Pacific economies.

Abstract

The Nelson-Siegel, Svensson, and De Rezende-Ferreira models are the most common approaches currently used for modeling the term structures of the risk-free interest rates. However, when markets become turbulent, they may not provide reliable results. Given the importance of the term structure in determining the time value of money and thus in pricing all securities, this study aims to improve the statistical performance of the above mentioned models using more advanced approaches, starting from evolutionary algorithms, such as genetic algorithms and particle swarm optimization, to a more comprehensive hybrid approach, combining the above mentioned artificial intelligence-based methodologies with the traditional Levenberg-Marquardt method. When also this latter approach is unsatisfactory, we rely on machine learning techniques, specifically using Gaussian Process Regression. This study considers the currencies of eight countries belonging to the Pacific area, in addition to the US dollar, as a reference currency. Results show that the use of artificial intelligence-based approaches improve the performance of traditional parametric models currently used in the field.

2024

Creating a More Prudent Value-at-Risk of an Options Portfolio Selecting Optimal Market Variables

Abstract

In this study, we have designed an advanced tool for comprehensive analysis of the financial market with pronounced emphasis on option pricing and risk quantification of 11 different assets in a portfolio. The central part of the paper focused on volatility and drift estimation through a suite of methodologies followed by various option pricing models operationalized through Monte Carlo simulation techniques. Following this, we compute the Value at Risk (VaR) metrics which is an essential element in the domain of financial risk management and finally compute the quantile functions to assess the potential financial losses in cases where the markets go through a downturn.

Demand Forecasting for a Manufacturing Company

Abstract

This study aims to forecast future demand for a manufacturing company in the automotive sector by leveraging a comprehensive sales dataset spanning from 2012 to 2024. The research begins by analyzing the historical sales patterns of all products, classifying their time series behavior to group them into distinct families based on their demand characteristics. This classification enables the selection of the most suitable forecasting model for each product category, ensuring tailored and accurate predictions.The methodology explores multiple forecasting approaches, including SARIMA, recurrent neural networks (RNNs), and specialized intermittent demand models like TSB, acknowledging that no single model universally outperforms others. Instead, the optimal choice depends on the unique patterns and features of each product’s historical demand data.