Forecasting in Industry andFinancial Context

Abstract Chapter 1 The first chapter of my PhD thesis addresses the demand forecasting problem for VHIT, the company sponsoring my doctoral research. The objective of the thesis is to predict the future value of the sales for the different products sold by the company. Forecasting in the manufacturing sector is considered particularly complex, as the time series of sold products are often intermittent and therefore difficult to forecast. These types of time series cannot be accurately predicted using traditional models from econometrics or machine learning but require forecasting methods specifically designed for intermittent time series. In our case, VHIT sells hundreds of different products, and the majority of them are characterized by intermittent behavior; however, there are also products with more linear time series. Therefore, what we did was to create a forecasting engine able to classify the different types of products and then select the most suitable forecasting model for each type. The results obtained show that, generally, products with intermittent time series are predicted with greater accuracy than more linear ones. This contradicts what is stated in the literature but is justified by the fact that in the case of VHIT, intermittent and linear time series are not equally distributed. In fact, we have a minority of non-intermittent time series that account for the vast majority of revenue, while we also have a large number of products with intermittent time series that, in terms of revenue, have little value. These products record very few sporadic orders or even none at all, making them easier to predict. Abstract Chapter 2 The Nelson-Siegel, Svensson, and De Rezende-Ferreira-Ferreira models represent some of the most widely adopted parametric frameworks for modeling the term structure of risk-free interest rates. Nonetheless, their reliability may diminish under volatile market conditions. Given the central role of the term structure in determining the time value of money and in pricing financial instruments, this study seeks to enhance the statistical robustness of these models by integrating nature-inspired global optimization algorithms, specifically Genetic Algorithms (GA), Particle Swarm Optimization (PSO), and Simulated Annealing (SA). These techniques are employed to explore the search space and identify promising regions, which are then used to initialize local optimization routines, improving convergence and reducing the risk to fall in a local minima. Finally, to validate our results against a challenging benchmark, we compare the performance of the parametric models with those obtained using a non-parametric machine learning approach, namely Gaussian Process Regression. The analysis was conducted on three distinct datasets to assess the robustness of the results: the first includes currencies from five developed economies, the second comprises nine currencies from Pacific countries, and the third consists of sovereign yield curves from five European nations. The findings confirm that integrating nature-inspired global optimization algorithms substantially enhances the fitting performance of traditional parametric models, making them competitive with machine learning approaches a comparison that was previously unimaginable in this context. Abstract Chapter 3 The Value at Risk (VaR) is one of the most widely used methods in the financial sector for assessing the risk associated with a portfolio over a specified period and is generally applied to portfolios composed of linear assets, where risk is more straightforward to quantify. In our study, aiming to provide an innovative contribution, we focus on nonlinear instruments such as options, which exhibit significant nonlinearity and require a different approach to risk estimation. Our objective is to calculate the most conservative Value at Risk for an option portfolio over a future time horizon, leveraging the Monte Carlo method to project the portfolio’s performance and ensure a robust VaR estimation. To outline the most pessimistic scenario, we analyze key factors influencing the valuation of options, specifically volatility, drift, and dividends, using multiple estimation techniques based on historical data. By extracting only the extreme scenarios: ”maximum and minimum values across different models” we ensure a comprehensive evaluation of risk exposure, performing extensive simulations to derive the worst-case VaR for the portfolio under varying market conditions.