|
on Forecasting |
| By: | Tae-Hwy Lee; Ekaterina Seregina |
| Abstract: | In this paper we develop a novel method of combining many forecasts based on a machine learning algorithm called Graphical LASSO. We visualize forecast errors from different forecasters as a network of interacting entities and generalize network inference in the presence of common factor structure and structural breaks. First, we note that forecasters often use common information and hence make common mistakes, which makes the forecast errors exhibit common factor structures. We propose the Factor Graphical LASSO (Factor GLASSO), which separates common forecast errors from the idiosyncratic errors and exploits sparsity of the precision matrix of the latter. Second, since the network of experts changes over time as a response to unstable environments such as recessions, it is unreasonable to assume constant forecast combination weights. Hence, we propose Regime-Dependent Factor Graphical LASSO (RD-Factor GLASSO) and develop its scalable implementation using the Alternating Direction Method of Multipliers (ADMM) to estimate regime-dependent forecast combination weights. The empirical application to forecasting macroeconomic series using the data of the European Central Bank's Survey of Professional Forecasters (ECB SPF) demonstrates superior performance of a combined forecast using Factor GLASSO and RD-Factor GLASSO. |
| Date: | 2022–09 |
| URL: | http://d.repec.org/n?u=RePEc:arx:papers:2209.01697&r= |
| By: | Prüser, Jan; Blagov, Boris |
| Abstract: | We propose a prior for VAR models that exploits the panel structure of macroeconomic time series while also providing shrinkage towards zero to address overfitting concerns. The prior is flexible as it detects shared dynamics of individual variables across endogenously determined groups of countries. We demonstrate the usefulness of our approach via a Monte Carlo study and use our model to capture the hidden homo- and heterogeneities of the euro area member states. Combining pairwise pooling with zero shrinkage delivers sharper parameter inference that improves point and density forecasts over only zero shrinkage or only pooling specifications, and helps with structural analysis by lowering the estimation uncertainty. |
| Keywords: | BVAR,shrinkage,forecasting,structural analysis |
| JEL: | C11 C32 C53 E37 |
| Date: | 2022 |
| URL: | http://d.repec.org/n?u=RePEc:zbw:rwirep:960&r= |
| By: | du Plessis, Emile; Fritsche, Ulrich |
| Abstract: | A reflection on the lackluster growth over the decade since the Global Financial Crisis has renewed interest in preventative measures for a long-standing problem. Advances in machine learning algorithms during this period present promising forecasting solutions. In this context, the paper develops new forecasting methods for an old problem by employing 13 machine learning algorithms to study 147 year of systemic financial crises across 17 countries. It entails 12 leading indicators comprising real, banking and external sectors. Four modelling dimensions encompassing a contemporaneous pooled format through an expanding window, transformations with a lag structure and 20-year rolling window as well as individual format are implemented to assess performance through recursive out-of-sample forecasts. Findings suggest fixed capital formation is the most important variable. GDP per capita and consumer inflation have increased in prominence whereas debt-to-GDP, stock market and consumption were dominant at the turn of the 20th century. Through a lag structure, banking sector predictors on average describe 28 percent of the variation in crisis prevalence, real sector 64 percent and external sector 8 percent. A lag structure and rolling window both improve on optimised contemporaneous and individual country formats. Nearly half of all algorithms reach peak performance through a lag structure. As measured through AUC, F1 and Brier scores, top performing machine learning methods consistently produce high accuracy rates, with both random forests and gradient boosting in front with 77 percent correct forecasts. Top models contribute added value above 20 percentage points in most instances and deals with a high degree of complexity across several countries. |
| Keywords: | machine learning,systemic financial crises,leading indicators,forecasting,early warning signal |
| JEL: | C14 C15 C32 C35 C53 E37 E44 G21 |
| Date: | 2022 |
| URL: | http://d.repec.org/n?u=RePEc:zbw:uhhwps:67&r= |
| By: | Sayar Karmakar (Department of Statistics, University of Florida, 230 Newell Drive, Gainesville, FL, 32601, USA); Rangan Gupta (Department of Economics, University of Pretoria, Private Bag X20, Hatfield 0028, South Africa); Oguzhan Cepni (Copenhagen Business School, Department of Economics, Porcel16A, Frederiksberg DK-2000, Denmark; Central Bank of the Republic of Turkey, Haci Bayram Mah. Istiklal Cad. No:10 06050, Ankara, Turkey); Lavinia Rognone (Alliance Manchester Business School, The University of Manchester, Booth St W, Manchester M15 6PB, UK) |
| Abstract: | We investigate the ability of textual analysis-based metrics of physical or transition risks associated with climate change in forecasting the daily volume of trade contracts of gold. Given the count-valued nature of gold volume data, our econometric framework is a loglinear Poisson integer-valued generalized autoregressive conditional heteroskedasticity (INGARCH) model with a particular climate change-related covariate. We detect a significant predictive power for gold volume at 5- and 22-day-ahead horizons when we extend our model using physical risks. Given the underlying positively evolving impact of such risks on the trading volume of gold, as derived from a full-sample analysis using a time-varying INGARCH model, we can say that gold acts as a hedge against physical risks at 1-week and 1-month horizons. Such a characteristic is also detected for platinum, and to a lesser extent, for palladium, but not silver. Our results have important investment implications. |
| Keywords: | Climate Risks, Precious Metals, Forecasting, Trading Volumes, Count Data, INGARCH |
| JEL: | C22 C53 Q02 Q54 |
| Date: | 2022–09 |
| URL: | http://d.repec.org/n?u=RePEc:pre:wpaper:202241&r= |
| By: | James Younker |
| Abstract: | Forecast combinations, also known as ensemble models, routinely require practitioners to select a model from a massive number of potential candidates. Ten explanatory variables can be grouped into 2^1078 forecast combinations, and the number of possibilities increases further to 2^(1078+2^1078) if we allow for forecast combinations of forecast combinations. This paper derives a calculation for the effective degrees of freedom of a forecast combination under a set of general conditions for linear models. It also supports this calculation with simulations. The result allows users to perform several other computations, including the F-test and various information criteria. These computations are particularly useful when there are too many candidate models to evaluate out of sample. Furthermore, computing effective degrees of freedom shows that the complexity cost of a forecast combination is driven by the parameters in the weighting scheme and the weighted average of parameters in the auxiliary models as opposed to the number of auxiliary models. This identification of complexity cost contributions can help practitioners make informed choices about forecast combination design. |
| Keywords: | Econometric and statistical methods |
| JEL: | C01 C02 C1 C13 C5 C50 C51 C52 C53 |
| Date: | 2022–09 |
| URL: | http://d.repec.org/n?u=RePEc:bca:bocadp:22-19&r= |
| By: | Lena Sasal; Tanujit Chakraborty; Abdenour Hadid |
| Abstract: | Deep learning utilizing transformers has recently achieved a lot of success in many vital areas such as natural language processing, computer vision, anomaly detection, and recommendation systems, among many others. Among several merits of transformers, the ability to capture long-range temporal dependencies and interactions is desirable for time series forecasting, leading to its progress in various time series applications. In this paper, we build a transformer model for non-stationary time series. The problem is challenging yet crucially important. We present a novel framework for univariate time series representation learning based on the wavelet-based transformer encoder architecture and call it W-Transformer. The proposed W-Transformers utilize a maximal overlap discrete wavelet transformation (MODWT) to the time series data and build local transformers on the decomposed datasets to vividly capture the nonstationarity and long-range nonlinear dependencies in the time series. Evaluating our framework on several publicly available benchmark time series datasets from various domains and with diverse characteristics, we demonstrate that it performs, on average, significantly better than the baseline forecasters for short-term and long-term forecasting, even for datasets that consist of only a few hundred training samples. |
| Date: | 2022–09 |
| URL: | http://d.repec.org/n?u=RePEc:arx:papers:2209.03945&r= |
| By: | Aurelio Tobias (Spanish Research Council (CSIC), Barcelona, Spain); Ben Armstrong (Spanish Research Council (CSIC), Barcelona, Spain); Antonio Gasparrini (Spanish Research Council (CSIC), Barcelona, Spain) |
| Abstract: | The distributed lag non-linear models (DLNMs) represent a modelling framework to flexibly describe associations showing potentially non-linear and delayed effects in time-series data. This methodology rests on the definition of a crossbasis, a bi-dimensional functional space combining two sets of basis functions, which specify the relationships in the dimensions of predictor and lags, respectively. DLNMs have been widely used in environmental epidemiology to investigate the short-term associations between environmental exposures, such as weather variables or air pollution, and health outcomes, such as mortality counts or disease-specific hospital admissions. We implemented the DLNMs framework in Stata through the crossbasis command to generate the basis variables that can be fitted in a broad range of regression models. In addition, the post estimation commands crossbgraph and crossbslices allow interpreting the results, emphasizing graphical representation, after the regression model fit. We present an overview of the capabilities of these new user-developed commands and describe the practical steps to fit and interpret DLNMs with an example of real data to represent the relationship between temperature and mortality in London during the period 2002-2006. |
| Date: | 2022–09–10 |
| URL: | http://d.repec.org/n?u=RePEc:boc:lsug22:09&r= |