Forecasting economic and financial time series with non-linear models

In this paper we discuss the current state-of-the-art in estimating, evaluating, and selecting among non-linear forecasting models for economic and financial time series. We review theoretical and empirical issues, including predictive density, interval and point evaluation and model selection, loss functions, data-mining, and aggregation. In addition, we argue that although the evidence in favor of constructing forecasts using non-linear models is rather sparse, there is reason to be optimistic. However, much remains to be done. Finally, we outline a variety of topics for future research, and discuss a number of areas which have received considerable attention in the recent literature, but where many questions remain.

Statistical models for short-term wind power ramp forecasting

La predicción de energía eólica ha desempeñado en la última década un papel fundamental en el aprovechamiento de este recurso renovable, ya que permite reducir el impacto que tiene la naturaleza fluctuante del viento en la actividad de diversos agentes implicados en su integración, tales como el operador del sistema o los agentes del mercado eléctrico. Los altos niveles de penetración eólica alcanzados recientemente por algunos países han puesto de manifiesto la necesidad de mejorar las predicciones durante eventos en los que se experimenta una variación importante de la potencia generada por un parque o un conjunto de ellos en un tiempo relativamente corto (del orden de unas pocas horas). Estos eventos, conocidos como rampas, no tienen una única causa, ya que pueden estar motivados por procesos meteorológicos que se dan en muy diferentes escalas espacio-temporales, desde el paso de grandes frentes en la macroescala a procesos convectivos locales como tormentas. Además, el propio proceso de conversión del viento en energía eléctrica juega un papel relevante en la ocurrencia de rampas debido, entre otros factores, a la relación no lineal que impone la curva de potencia del aerogenerador, la desalineación de la máquina con respecto al viento y la interacción aerodinámica entre aerogeneradores. En este trabajo se aborda la aplicación de modelos estadísticos a la predicción de rampas a muy corto plazo. Además, se investiga la relación de este tipo de eventos con procesos atmosféricos en la macroescala. Los modelos se emplean para generar predicciones de punto a partir del modelado estocástico de una serie temporal de potencia generada por un parque eólico. Los horizontes de predicción considerados van de una a seis horas. Como primer paso, se ha elaborado una metodología para caracterizar rampas en series temporales. La denominada función-rampa está basada en la transformada wavelet y proporciona un índice en cada paso temporal. Este índice caracteriza la intensidad de rampa en base a los gradientes de potencia experimentados en un rango determinado de escalas temporales. Se han implementado tres tipos de modelos predictivos de cara a evaluar el papel que juega la complejidad de un modelo en su desempeño: modelos lineales autorregresivos (AR), modelos de coeficientes variables (VCMs) y modelos basado en redes neuronales (ANNs). Los modelos se han entrenado en base a la minimización del error cuadrático medio y la configuración de cada uno de ellos se ha determinado mediante validación cruzada. De cara a analizar la contribución del estado macroescalar de la atmósfera en la predicción de rampas, se ha propuesto una metodología que permite extraer, a partir de las salidas de modelos meteorológicos, información relevante para explicar la ocurrencia de estos eventos. La metodología se basa en el análisis de componentes principales (PCA) para la síntesis de la datos de la atmósfera y en el uso de la información mutua (MI) para estimar la dependencia no lineal entre dos señales. Esta metodología se ha aplicado a datos de reanálisis generados con un modelo de circulación general (GCM) de cara a generar variables exógenas que posteriormente se han introducido en los modelos predictivos. Los casos de estudio considerados corresponden a dos parques eólicos ubicados en España. Los resultados muestran que el modelado de la serie de potencias permitió una mejora notable con respecto al modelo predictivo de referencia (la persistencia) y que al añadir información de la macroescala se obtuvieron mejoras adicionales del mismo orden. Estas mejoras resultaron mayores para el caso de rampas de bajada. Los resultados también indican distintos grados de conexión entre la macroescala y la ocurrencia de rampas en los dos parques considerados. Abstract One of the main drawbacks of wind energy is that it exhibits intermittent generation greatly depending on environmental conditions. Wind power forecasting has proven to be an effective tool for facilitating wind power integration from both the technical and the economical perspective. Indeed, system operators and energy traders benefit from the use of forecasting techniques, because the reduction of the inherent uncertainty of wind power allows them the adoption of optimal decisions. Wind power integration imposes new challenges as higher wind penetration levels are attained. Wind power ramp forecasting is an example of such a recent topic of interest. The term ramp makes reference to a large and rapid variation (1-4 hours) observed in the wind power output of a wind farm or portfolio. Ramp events can be motivated by a broad number of meteorological processes that occur at different time/spatial scales, from the passage of large-scale frontal systems to local processes such as thunderstorms and thermally-driven flows. Ramp events may also be conditioned by features related to the wind-to-power conversion process, such as yaw misalignment, the wind turbine shut-down and the aerodynamic interaction between wind turbines of a wind farm (wake effect). This work is devoted to wind power ramp forecasting, with special focus on the connection between the global scale and ramp events observed at the wind farm level. The framework of this study is the point-forecasting approach. Time series based models were implemented for very short-term prediction, this being characterised by prediction horizons up to six hours ahead. As a first step, a methodology to characterise ramps within a wind power time series was proposed. The so-called ramp function is based on the wavelet transform and it provides a continuous index related to the ramp intensity at each time step. The underlying idea is that ramps are characterised by high power output gradients evaluated under different time scales. A number of state-of-the-art time series based models were considered, namely linear autoregressive (AR) models, varying-coefficient models (VCMs) and artificial neural networks (ANNs). This allowed us to gain insights into how the complexity of the model contributes to the accuracy of the wind power time series modelling. The models were trained in base of a mean squared error criterion and the final set-up of each model was determined through cross-validation techniques. In order to investigate the contribution of the global scale into wind power ramp forecasting, a methodological proposal to identify features in atmospheric raw data that are relevant for explaining wind power ramp events was presented. The proposed methodology is based on two techniques: principal component analysis (PCA) for atmospheric data compression and mutual information (MI) for assessing non-linear dependence between variables. The methodology was applied to reanalysis data generated with a general circulation model (GCM). This allowed for the elaboration of explanatory variables meaningful for ramp forecasting that were utilized as exogenous variables by the forecasting models. The study covered two wind farms located in Spain. All the models outperformed the reference model (the persistence) during both ramp and non-ramp situations. Adding atmospheric information had a noticeable impact on the forecasting performance, specially during ramp-down events. Results also suggested different levels of connection between the ramp occurrence at the wind farm level and the global scale.

A hybrid commodity price-forecasting model applied to the sugar-alcohol sector

Accurate price forecasting for agricultural commodities can have significant decision-making implications for suppliers, especially those of biofuels, where the agriculture and energy sectors intersect. Environmental pressures and high oil prices affect demand for biofuels and have reignited the discussion about effects on food prices. Suppliers in the sugar-alcohol sector need to decide the ideal proportion of ethanol and sugar to optimise their financial strategy. Prices can be affected by exogenous factors, such as exchange rates and interest rates, as well as non-observable variables like the convenience yield, which is related to supply shortages. The literature generally uses two approaches: artificial neural networks (ANNs), which are recognised as being in the forefront of exogenous-variable analysis, and stochastic models such as the Kalman filter, which is able to account for non-observable variables. This article proposes a hybrid model for forecasting the prices of agricultural commodities that is built upon both approaches and is applied to forecast the price of sugar. The Kalman filter considers the structure of the stochastic process that describes the evolution of prices. Neural networks allow variables that can impact asset prices in an indirect, nonlinear way, what cannot be incorporated easily into traditional econometric models.

Short-term Load Forecasting Based on Load Profiling

Load forecasting has gradually becoming a major field of research in electricity industry. Therefore, Load forecasting is extremely important for the electric sector under deregulated environment as it provides a useful support to the power system management. Accurate power load forecasting models are required to the operation and planning of a utility company, and they have received increasing attention from researches of this field study. Many mathematical methods have been developed for load forecasting. This work aims to develop and implement a load forecasting method for short-term load forecasting (STLF), based on Holt-Winters exponential smoothing and an artificial neural network (ANN). One of the main contributions of this paper is the application of Holt-Winters exponential smoothing approach to the forecasting problem and, as an evaluation of the past forecasting work, data mining techniques are also applied to short-term Load forecasting. Both ANN and Holt-Winters exponential smoothing approaches are compared and evaluated.

Short-term Wind Speed Forecasting using Support Vector Machines

Wind speed forecasting has been becoming an important field of research to support the electricity industry mainly due to the increasing use of distributed energy sources, largely based on renewable sources. This type of electricity generation is highly dependent on the weather conditions variability, particularly the variability of the wind speed. Therefore, accurate wind power forecasting models are required to the operation and planning of wind plants and power systems. A Support Vector Machines (SVM) model for short-term wind speed is proposed and its performance is evaluated and compared with several artificial neural network (ANN) based approaches. A case study based on a real database regarding 3 years for predicting wind speed at 5 minutes intervals is presented.

Forecasting human entrances at a commercial store using facial recognition data

Dissertação de mestrado integrado em Engenharia e Gestão de Sistemas de Informação

Factor models, VARMA processes and parameter instability with applications in macroeconomics

Avec les avancements de la technologie de l'information, les données temporelles économiques et financières sont de plus en plus disponibles. Par contre, si les techniques standard de l'analyse des séries temporelles sont utilisées, une grande quantité d'information est accompagnée du problème de dimensionnalité. Puisque la majorité des séries d'intérêt sont hautement corrélées, leur dimension peut être réduite en utilisant l'analyse factorielle. Cette technique est de plus en plus populaire en sciences économiques depuis les années 90. Étant donnée la disponibilité des données et des avancements computationnels, plusieurs nouvelles questions se posent. Quels sont les effets et la transmission des chocs structurels dans un environnement riche en données? Est-ce que l'information contenue dans un grand ensemble d'indicateurs économiques peut aider à mieux identifier les chocs de politique monétaire, à l'égard des problèmes rencontrés dans les applications utilisant des modèles standards? Peut-on identifier les chocs financiers et mesurer leurs effets sur l'économie réelle? Peut-on améliorer la méthode factorielle existante et y incorporer une autre technique de réduction de dimension comme l'analyse VARMA? Est-ce que cela produit de meilleures prévisions des grands agrégats macroéconomiques et aide au niveau de l'analyse par fonctions de réponse impulsionnelles? Finalement, est-ce qu'on peut appliquer l'analyse factorielle au niveau des paramètres aléatoires? Par exemple, est-ce qu'il existe seulement un petit nombre de sources de l'instabilité temporelle des coefficients dans les modèles macroéconomiques empiriques? Ma thèse, en utilisant l'analyse factorielle structurelle et la modélisation VARMA, répond à ces questions à travers cinq articles. Les deux premiers chapitres étudient les effets des chocs monétaire et financier dans un environnement riche en données. Le troisième article propose une nouvelle méthode en combinant les modèles à facteurs et VARMA. Cette approche est appliquée dans le quatrième article pour mesurer les effets des chocs de crédit au Canada. La contribution du dernier chapitre est d'imposer la structure à facteurs sur les paramètres variant dans le temps et de montrer qu'il existe un petit nombre de sources de cette instabilité. Le premier article analyse la transmission de la politique monétaire au Canada en utilisant le modèle vectoriel autorégressif augmenté par facteurs (FAVAR). Les études antérieures basées sur les modèles VAR ont trouvé plusieurs anomalies empiriques suite à un choc de la politique monétaire. Nous estimons le modèle FAVAR en utilisant un grand nombre de séries macroéconomiques mensuelles et trimestrielles. Nous trouvons que l'information contenue dans les facteurs est importante pour bien identifier la transmission de la politique monétaire et elle aide à corriger les anomalies empiriques standards. Finalement, le cadre d'analyse FAVAR permet d'obtenir les fonctions de réponse impulsionnelles pour tous les indicateurs dans l'ensemble de données, produisant ainsi l'analyse la plus complète à ce jour des effets de la politique monétaire au Canada. Motivée par la dernière crise économique, la recherche sur le rôle du secteur financier a repris de l'importance. Dans le deuxième article nous examinons les effets et la propagation des chocs de crédit sur l'économie réelle en utilisant un grand ensemble d'indicateurs économiques et financiers dans le cadre d'un modèle à facteurs structurel. Nous trouvons qu'un choc de crédit augmente immédiatement les diffusions de crédit (credit spreads), diminue la valeur des bons de Trésor et cause une récession. Ces chocs ont un effet important sur des mesures d'activité réelle, indices de prix, indicateurs avancés et financiers. Contrairement aux autres études, notre procédure d'identification du choc structurel ne requiert pas de restrictions temporelles entre facteurs financiers et macroéconomiques. De plus, elle donne une interprétation des facteurs sans restreindre l'estimation de ceux-ci. Dans le troisième article nous étudions la relation entre les représentations VARMA et factorielle des processus vectoriels stochastiques, et proposons une nouvelle classe de modèles VARMA augmentés par facteurs (FAVARMA). Notre point de départ est de constater qu'en général les séries multivariées et facteurs associés ne peuvent simultanément suivre un processus VAR d'ordre fini. Nous montrons que le processus dynamique des facteurs, extraits comme combinaison linéaire des variables observées, est en général un VARMA et non pas un VAR comme c'est supposé ailleurs dans la littérature. Deuxièmement, nous montrons que même si les facteurs suivent un VAR d'ordre fini, cela implique une représentation VARMA pour les séries observées. Alors, nous proposons le cadre d'analyse FAVARMA combinant ces deux méthodes de réduction du nombre de paramètres. Le modèle est appliqué dans deux exercices de prévision en utilisant des données américaines et canadiennes de Boivin, Giannoni et Stevanovic (2010, 2009) respectivement. Les résultats montrent que la partie VARMA aide à mieux prévoir les importants agrégats macroéconomiques relativement aux modèles standards. Finalement, nous estimons les effets de choc monétaire en utilisant les données et le schéma d'identification de Bernanke, Boivin et Eliasz (2005). Notre modèle FAVARMA(2,1) avec six facteurs donne les résultats cohérents et précis des effets et de la transmission monétaire aux États-Unis. Contrairement au modèle FAVAR employé dans l'étude ultérieure où 510 coefficients VAR devaient être estimés, nous produisons les résultats semblables avec seulement 84 paramètres du processus dynamique des facteurs. L'objectif du quatrième article est d'identifier et mesurer les effets des chocs de crédit au Canada dans un environnement riche en données et en utilisant le modèle FAVARMA structurel. Dans le cadre théorique de l'accélérateur financier développé par Bernanke, Gertler et Gilchrist (1999), nous approximons la prime de financement extérieur par les credit spreads. D'un côté, nous trouvons qu'une augmentation non-anticipée de la prime de financement extérieur aux États-Unis génère une récession significative et persistante au Canada, accompagnée d'une hausse immédiate des credit spreads et taux d'intérêt canadiens. La composante commune semble capturer les dimensions importantes des fluctuations cycliques de l'économie canadienne. L'analyse par décomposition de la variance révèle que ce choc de crédit a un effet important sur différents secteurs d'activité réelle, indices de prix, indicateurs avancés et credit spreads. De l'autre côté, une hausse inattendue de la prime canadienne de financement extérieur ne cause pas d'effet significatif au Canada. Nous montrons que les effets des chocs de crédit au Canada sont essentiellement causés par les conditions globales, approximées ici par le marché américain. Finalement, étant donnée la procédure d'identification des chocs structurels, nous trouvons des facteurs interprétables économiquement. Le comportement des agents et de l'environnement économiques peut varier à travers le temps (ex. changements de stratégies de la politique monétaire, volatilité de chocs) induisant de l'instabilité des paramètres dans les modèles en forme réduite. Les modèles à paramètres variant dans le temps (TVP) standards supposent traditionnellement les processus stochastiques indépendants pour tous les TVPs. Dans cet article nous montrons que le nombre de sources de variabilité temporelle des coefficients est probablement très petit, et nous produisons la première évidence empirique connue dans les modèles macroéconomiques empiriques. L'approche Factor-TVP, proposée dans Stevanovic (2010), est appliquée dans le cadre d'un modèle VAR standard avec coefficients aléatoires (TVP-VAR). Nous trouvons qu'un seul facteur explique la majorité de la variabilité des coefficients VAR, tandis que les paramètres de la volatilité des chocs varient d'une façon indépendante. Le facteur commun est positivement corrélé avec le taux de chômage. La même analyse est faite avec les données incluant la récente crise financière. La procédure suggère maintenant deux facteurs et le comportement des coefficients présente un changement important depuis 2007. Finalement, la méthode est appliquée à un modèle TVP-FAVAR. Nous trouvons que seulement 5 facteurs dynamiques gouvernent l'instabilité temporelle dans presque 700 coefficients.

Real-time factor model forecasting and the effects of instability

Factor forecasting models are shown to deliver real-time gains over autoregressive models for US real activity variables during the recent period, but are less successful for nominal variables. The gains are largely due to the Financial Crisis period, and are primarily at the shortest (one quarter ahead) horizon. Excluding the pre-Great Moderation years from the factor forecasting model estimation period (but not from the data used to extract factors) results in a marked fillip in factor model forecast accuracy, but does the same for the AR model forecasts. The relative performance of the factor models compared to the AR models is largely unaffected by whether the exercise is in real time or is pseudo out-of-sample.

Forecasting Brazilian inflation by its aggregate and disaggregated data: a test of predictive power by forecast horizon

This work aims to compare the forecast efficiency of different types of methodologies applied to Brazilian Consumer inflation (IPCA). We will compare forecasting models using disaggregated and aggregated data over twelve months ahead. The disaggregated models were estimated by SARIMA and will have different levels of disaggregation. Aggregated models will be estimated by time series techniques such as SARIMA, state-space structural models and Markov-switching. The forecasting accuracy comparison will be made by the selection model procedure known as Model Confidence Set and by Diebold-Mariano procedure. We were able to find evidence of forecast accuracy gains in models using more disaggregated data

Analysis of NDVI time series using cross-correlation and forecasting methods for monitoring sugarcane fields in Brazil

Brazil is the largest sugarcane producer in the world and has a privileged position to attend to national and international market places. To maintain the high production of sugarcane, it is fundamental to improve the forecasting models of crop seasons through the use of alternative technologies, such as remote sensing. Thus, the main purpose of this article is to assess the results of two different statistical forecasting methods applied to an agroclimatic index (the water requirement satisfaction index; WRSI) and the sugarcane spectral response (normalized difference vegetation index; NDVI) registered on National Oceanic and Atmospheric Administration Advanced Very High Resolution Radiometer (NOAA-AVHRR) satellite images. We also evaluated the cross-correlation between these two indexes. According to the results obtained, there are meaningful correlations between NDVI and WRSI with time lags. Additionally, the adjusted model for NDVI presented more accurate results than the forecasting models for WRSI. Finally, the analyses indicate that NDVI is more predictable due to its seasonality and the WRSI values are more variable making it difficult to forecast.

Earthquake forecasting and seismic hazard analysis: some insights on the testing phase and the modeling

This thesis is divided in three chapters. In the first chapter we analyse the results of the world forecasting experiment run by the Collaboratory for the Study of Earthquake Predictability (CSEP). We take the opportunity of this experiment to contribute to the definition of a more robust and reliable statistical procedure to evaluate earthquake forecasting models. We first present the models and the target earthquakes to be forecast. Then we explain the consistency and comparison tests that are used in CSEP experiments to evaluate the performance of the models. Introducing a methodology to create ensemble forecasting models, we show that models, when properly combined, are almost always better performing that any single model. In the second chapter we discuss in depth one of the basic features of PSHA: the declustering of the seismicity rates. We first introduce the Cornell-McGuire method for PSHA and we present the different motivations that stand behind the need of declustering seismic catalogs. Using a theorem of the modern probability (Le Cam's theorem) we show that the declustering is not necessary to obtain a Poissonian behaviour of the exceedances that is usually considered fundamental to transform exceedance rates in exceedance probabilities in the PSHA framework. We present a method to correct PSHA for declustering, building a more realistic PSHA. In the last chapter we explore the methods that are commonly used to take into account the epistemic uncertainty in PSHA. The most widely used method is the logic tree that stands at the basis of the most advanced seismic hazard maps. We illustrate the probabilistic structure of the logic tree, and then we show that this structure is not adequate to describe the epistemic uncertainty. We then propose a new probabilistic framework based on the ensemble modelling that properly accounts for epistemic uncertainties in PSHA.

Electricity price forecasting accounting for renewable energies: optimal combined forecasts

Electricity price forecasting is an interesting problem for all the agents involved in electricity market operation. For instance, every profit maximisation strategy is based on the computation of accurate one-day-ahead forecasts, which is why electricity price forecasting has been a growing field of research in recent years. In addition, the increasing concern about environmental issues has led to a high penetration of renewable energies, particularly wind. In some European countries such as Spain, Germany and Denmark, renewable energy is having a deep impact on the local power markets. In this paper, we propose an optimal model from the perspective of forecasting accuracy, and it consists of a combination of several univariate and multivariate time series methods that account for the amount of energy produced with clean energies, particularly wind and hydro, which are the most relevant renewable energy sources in the Iberian Market. This market is used to illustrate the proposed methodology, as it is one of those markets in which wind power production is more relevant in terms of its percentage of the total demand, but of course our method can be applied to any other liberalised power market. As far as our contribution is concerned, first, the methodology proposed by García-Martos et al(2007 and 2012) is generalised twofold: we allow the incorporation of wind power production and hydro reservoirs, and we do not impose the restriction of using the same model for 24h. A computational experiment and a Design of Experiments (DOE) are performed for this purpose. Then, for those hours in which there are two or more models without statistically significant differences in terms of their forecasting accuracy, a combination of forecasts is proposed by weighting the best models(according to the DOE) and minimising the Mean Absolute Percentage Error (MAPE). The MAPE is the most popular accuracy metric for comparing electricity price forecasting models. We construct the combi nation of forecasts by solving several nonlinear optimisation problems that allow computation of the optimal weights for building the combination of forecasts. The results are obtained by a large computational experiment that entails calculating out-of-sample forecasts for every hour in every day in the period from January 2007 to Decem ber 2009. In addition, to reinforce the value of our methodology, we compare our results with those that appear in recent published works in the field. This comparison shows the superiority of our methodology in terms of forecasting accuracy.

Election Forecasting under Opaque Conditions

Election forecasting models assume retrospective economic voting and clear mechanisms of accountability. Previous research indeed indicates that incumbent political parties are being held accountable for the state of the economy. In this article we develop a ‘hard case’ for the assumptions of election forecasting models. Belgium is a multiparty system with perennial coalition governments. Furthermore, Belgium has two completely segregated party systems (Dutch and French language). Since the prime minister during the period 1974-2011 has always been a Dutch language politician, French language voters could not even vote for the prime minister, so this cognitive shortcut to establish political accountability is not available. Results of an analysis for the French speaking parties (1981-2010) show that even in these conditions of opaque accountability, retrospective economic voting occurs as election results respond to indicators with regard to GDP and unemployment levels. Party membership figures can be used to model the popularity function in election forecasting.

Inverse forecasting:a new approach for predictive modelling

In the last two decades there have been substantial developments in the mathematical theory of inverse optimization problems, and their applications have expanded greatly. In parallel, time series analysis and forecasting have become increasingly important in various fields of research such as data mining, economics, business, engineering, medicine, politics, and many others. Despite the large uses of linear programming in forecasting models there is no a single application of inverse optimization reported in the forecasting literature when the time series data is available. Thus the goal of this paper is to introduce inverse optimization into forecasting field, and to provide a streamlined approach to time series analysis and forecasting using inverse linear programming. An application has been used to demonstrate the use of inverse forecasting developed in this study. © 2007 Elsevier Ltd. All rights reserved.

Does money matter in inflation forecasting

This paper provides the most fully comprehensive evidence to date on whether or not monetary aggregates are valuable for forecasting US inflation in the early to mid 2000s. We explore a wide range of different definitions of money, including different methods of aggregation and different collections of included monetary assets. In our forecasting experiment we use two non-linear techniques, namely, recurrent neural networks and kernel recursive least squares regression - techniques that are new to macroeconomics. Recurrent neural networks operate with potentially unbounded input memory, while the kernel regression technique is a finite memory predictor. The two methodologies compete to find the best fitting US inflation forecasting models and are then compared to forecasts from a naive random walk model. The best models were non-linear autoregressive models based on kernel methods. Our findings do not provide much support for the usefulness of monetary aggregates in forecasting inflation.