|
on Forecasting |
| By: | Bartosz Uniejewski |
| Abstract: | The most commonly used form of regularization typically involves defining the penalty function as a L1 or L2 norm. However, numerous alternative approaches remain untested in practical applications. In this study, we apply ten different penalty functions to predict electricity prices and evaluate their performance under two different model structures and in two distinct electricity markets. The study reveals that LQ and elastic net consistently produce more accurate forecasts compared to other regularization types. In particular, they were the only types of penalty functions that consistently produced more accurate forecasts than the most commonly used LASSO. Furthermore, the results suggest that cross-validation outperforms Bayesian information criteria for parameter optimization, and performs as well as models with ex-post parameter selection. |
| Date: | 2024–04 |
| URL: | http://d.repec.org/n?u=RePEc:arx:papers:2404.03968&r=for |
| By: | Arkadiusz Lipiecki; Bartosz Uniejewski; Rafa{\l} Weron |
| Abstract: | Operational decisions relying on predictive distributions of electricity prices can result in significantly higher profits compared to those based solely on point forecasts. However, the majority of models developed in both academic and industrial settings provide only point predictions. To address this, we examine three postprocessing methods for converting point forecasts into probabilistic ones: Quantile Regression Averaging, Conformal Prediction, and the recently introduced Isotonic Distributional Regression. We find that while IDR demonstrates the most varied performance, combining its predictive distributions with those of the other two methods results in an improvement of ca. 7.5% compared to a benchmark model with normally distributed errors, over a 4.5-year test period in the German power market spanning the COVID pandemic and the war in Ukraine. Remarkably, the performance of this combination is at par with state-of-the-art Distributional Deep Neural Networks. |
| Date: | 2024–04 |
| URL: | http://d.repec.org/n?u=RePEc:arx:papers:2404.02270&r=for |
| By: | Pedro Afonso Fernandes |
| Abstract: | Economic forecasting is concerned with the estimation of some variable like gross domestic product (GDP) in the next period given a set of variables that describes the current situation or state of the economy, including industrial production, retail trade turnover or economic confidence. Neuro-dynamic programming (NDP) provides tools to deal with forecasting and other sequential problems with such high-dimensional states spaces. Whereas conventional forecasting methods penalises the difference (or loss) between predicted and actual outcomes, NDP favours the difference between temporally successive predictions, following an interactive and trial-and-error approach. Past data provides a guidance to train the models, but in a different way from ordinary least squares (OLS) and other supervised learning methods, signalling the adjustment costs between sequential states. We found that it is possible to train a GDP forecasting model with data concerned with other countries that performs better than models trained with past data from the tested country (Portugal). In addition, we found that non-linear architectures to approximate the value function of a sequential problem, namely, neural networks can perform better than a simple linear architecture, lowering the out-of-sample mean absolute forecast error (MAE) by 32% from an OLS model. |
| Date: | 2024–04 |
| URL: | http://d.repec.org/n?u=RePEc:arx:papers:2404.03737&r=for |