Tesis:
Modelos de predicción con factores inobservables en series temporales multivariantes. Aplicación a los precios del mercado eléctrico español.
- Autor: GARCIA MARTOS, Carolina
- Título: Modelos de predicción con factores inobservables en series temporales multivariantes. Aplicación a los precios del mercado eléctrico español.
- Fecha: 2010
- Materia: Sin materia definida
- Escuela: E.T.S. DE INGENIEROS INDUSTRIALES
- Departamentos: INGENIERIA DE ORGANIZACION, ADMINISTRACION DE EMPRESAS Y ESTADISTICA
- Acceso electrónico:
- Director/a 1º: RODRIGUEZ PUERTA, Julio
- Director/a 2º: SANCHEZ NARANJO, María Jesús
- Resumen: This work is framed within modelling and forecasting high-dimensional vectors of time series. Although detailed conslusions are provided at the end of each Chapter, a brief summary of them is included here, just before focusing on further lines of research that arise from this PhD thesis. The contributions may be summarized as follows: Development and implementation of univariate mixed models for electricity prices which improved, in terms of prediction errors, the performance of very cited work such as Conejo et al. (2005), Nogales et al. (2002), and Contreras at al. (2003). Each of the 24 hourly time series are modelled using a separate model, i.e., appliying the parallel (or periodic) approach where model and parameters depend on the hour of the day. In the evaluation, the out-of-sample forecasts have been evaluated for a long consecutive period, which is uncommon in literature. Development of a relatively simple Dynamic Factor Model based on the work by Pena and Box (1987) and Lee and Carter (1992) which improves the results of the univariate mixed model briefly described in 1. These methodologies could be of application to prices of any other power market. In Chapter 2, we briefly presented some results for the Ontario Market. 3. Development of the Seasonal Dynamic Factor Analysis (SeaDFA), which allows for dimension reduction in time series with seasonal behaviour. Seasonal dynamics are not often handled in empirical analyses, and most dynamic factor models assume stationarity of the factors and original series. 4. Proposal of a new bootstrap-based procedure, which is a successful alternative to the existing methods. Moreover, it does not need a backwards representation of the model. It is validated using a Monte Carlo simulation experiment and also applied to real data, and long-term forecasting intervals have been computed for electricity prices. This kind of forecasts is useful for managing adequately the risk present in bilateral contracts. The models developed for calculating short-term predictions are not accurate when using them for computing long-run forecasts. In Chapter 4 the possibility of conditionally heteroskedastic seasonal dynamic common factors is introduced. To date, dimensionality reduction techniques were adequate only for vectors of series with structure in conditional mean or in conditional variance, but not in both. A successful application to real data has been carried out, computing short and long term forecasts of electricity prices and their volatilities, as well as extracting dynamic common factors as well as common volatility factors. In Chapter 5 a decomposition technique for vectors of series into common and specific dynamics is provided. This methodology is applied to several European power market data, extracting common and specific dynamic features of the prices under study. These new models proposed (SeaDFA, Conditionally Heteroskedastic SeaDFA), the new bootstrap procedure as well as the decomposition technique presented in Chapter 5 can be of application to any high dimensional vector of series which present seasonality, structure both in conditional mean and variance, or when we are interested in extracting common and specific dynamic features from a panel of series. Examples of this kind of data are sets of macroeconomic variables, demographic and metheorological data or other data coming from electricity markets, such as load, among others.