|
on Forecasting |
| By: | Luigi Longo (IMT School for advanced studies); Massimo Riccaboni (IMT School for advanced studies); Armando Rungi (IMT School for advanced studies) |
| Abstract: | We propose an ensemble learning methodology to forecast the future US GDP growth release. Our approach combines a Recurrent Neural Network (RNN) with a Dynamic Factor model accounting for time-variation in mean with a General- ized Autoregressive Score (DFM-GAS). The analysis is based on a set of predictors encompassing a wide range of variables measured at different frequencies. The forecast exercise is aimed at evaluating the predictive ability of each model's com- ponent of the ensemble by considering variations in mean, potentially caused by recessions affecting the economy. Thus, we show how the combination of RNN and DFM-GAS improves forecasts of the US GDP growth rate in the aftermath of the 2008-09 global financial crisis. We find that a neural network ensemble markedly reduces the root mean squared error for the short-term forecast horizon. |
| Keywords: | macroeconomic forecasting; machine learning; neural networks; dynamic factor model; Covid-19 crisis |
| JEL: | C53 E37 |
| Date: | 2021–03 |
| URL: | http://d.repec.org/n?u=RePEc:ial:wpaper:2/2021&r=all |
| By: | Ly, Racine; Traore, Fousseini; Dia, Khadim |
| Abstract: | This paper applies a recurrent neural network (RNN) method to forecast cotton and oil prices. We show how these new tools from machine learning, particularly Long-Short Term Memory (LSTM) models, complement traditional methods. Our results show that machine learning methods fit reasonably well with the data but do not outperform systematically classical methods such as Autoregressive Integrated Moving Average (ARIMA) or the naïve models in terms of out of sample forecasts. However, averaging the forecasts from the two type of models provide better results compared to either method. Compared to the ARIMA and the LSTM, the Root Mean Squared Error (RMSE) of the average forecast was 0.21 and 21.49 percent lower, respectively, for cotton. For oil, the forecast averaging does not provide improvements in terms of RMSE. We suggest using a forecast averaging method and extending our analysis to a wide range of commodity prices. |
| Keywords: | WORLD; forecasting; models; prices; commodities; machine learning; neural networks; cotton; oils; Recurrent Neural networks; LSTM; commodity prices; Long-Short Term Memory; Autoregressive Integrated Moving Average (ARIMA) |
| Date: | 2021 |
| URL: | http://d.repec.org/n?u=RePEc:fpr:ifprid:2000&r=all |
| By: | Christian Conrad (Department of Economics, Heidelberg University, Germany; KOF Swiss Economic Institute, Switzerland; Rimini Centre for Economic Analysis); Robert F. Engle (New York University, Stern School of Business, USA; Rimini Centre for Economic Analysis) |
| Abstract: | We suggest a multiplicative factor multi frequency component GARCH model which exploits the empirical fact that the daily standardized forecast errors of standard GARCH models behave counter-cyclical when averaged at a lower frequency. For the new model, we derive the unconditional variance of the returns, the news impact function and multi-step-ahead volatility forecasts. We apply the model to the S&P 500, the FTSE 100 and the Hang Seng Index. We show that the long-term component of stock market volatility is driven by news about the macroeconomic outlook and monetary policy as well as policy-related news. The new component model significantly outperforms the nested one-component (GJR) GARCH and several HAR-type models in terms of out-of-sample forecasting. |
| Keywords: | Volatility forecasting, long- and short-term volatility, mixed frequency data, volatility cycles |
| JEL: | C53 C58 G12 |
| Date: | 2021–03 |
| URL: | http://d.repec.org/n?u=RePEc:rim:rimwps:21-05&r=all |
| By: | Matteo Garzoli (University of Lugano); Alberto Plazzi (Swiss Finance Institute; Universita' della Svizzera italiana); Rossen I. Valkanov (University of California, San Diego (UCSD) - Rady School of Management) |
| Abstract: | The Case-Shiller is the reference repeat-sales index for the U.S. residential real estate market, yet it is released with a two-month delay. We find that incorporating recent information from 71 financial and macro predictors improves backcasts, now-casts, and short-term forecasts of the index returns. Combining individual forecasts with recently-proposed weighting schemes delivers large improvements in forecast accuracy at all horizons. Additional gains obtain with mixed-data sampling methods that exploit the daily frequency of financial variables, reducing the mean square forecast error by as much as 13% compared to a simple autoregressive benchmark. The forecast improvements are largest during economic turmoils, throughout the 2020 COVID-19 pandemic period, and in more populous metropolitan areas. |
| Keywords: | Real estate, Case-Shiller, MIDAS, Forecasting, Big Data |
| JEL: | C22 C53 R30 |
| Date: | 2021–03 |
| URL: | http://d.repec.org/n?u=RePEc:chf:rpseri:rp2121&r=all |
| By: | Uddin, Gazi Salah (Asian Development Bank Institute); Tang, Ou (Asian Development Bank Institute); Sahamkhadam, Maziar (Asian Development Bank Institute); Taghizadeh-Hesary, Farhad (Asian Development Bank Institute); Yahya, Muhammad (Asian Development Bank Institute); Cerin, Pontus (Asian Development Bank Institute); Rehme, Jakob (Asian Development Bank Institute) |
| Abstract: | Short-term electricity price forecasting has received considerable attention in recent years. Despite this increased interest, the literature lacks a concrete consensus on the most suitable forecasting approach. We conduct an extensive empirical analysis to evaluate the short-term price forecasting dynamics of different regions in the Swedish electricity market (SEM). We utilized several forecasting approaches ranging from standard conditional volatility models to wavelet-based forecasting. In addition, we performed out-of-sample forecasting and back-testing, and we evaluated the performance of these models. Our empirical analysis indicates that an ARMA-GARCH framework with the student’s t-distribution significantly outperforms other frameworks. We only performed wavelet-based forecasting based on the MAPE. The results of the robust forecasting methods are capable of displaying the importance of proper forecasting process design, policy implications for market efficiency, and predictability in the SEM. |
| Keywords: | forecasting; Swedish electricity market; GARCH modeling; multi-scale analysis |
| JEL: | C53 G17 |
| Date: | 2021–01–13 |
| URL: | http://d.repec.org/n?u=RePEc:ris:adbiwp:1212&r=all |
| By: | Gary Koop; Stuart McIntyre; James Mitchell; Aubrey Poon |
| Abstract: | Expenditure side and income side GDP are both measured at the quarterly frequency in the US and contain measurement error. They are noisy proxies of `true’ GDP. Several econometric methods exist for producing estimates of true GDP which reconcile these noisy estimates. Recently, the authors of this paper developed a mixed frequency reconciliation model which produces monthly estimates of true GDP. In the present paper, we investigate whether this model continues to work well in the face of the extreme observations that occurred during the pandemic year of 2020 and consider several extensions of it. These extensions include stochastic volatility and error distributions that are fat tailed or explicitly allow for outliers. We also investigate the performance of conditional forecasting, where we estimate our models using data through 2019 and then use these to nowcast throughout 2020. Nowcasts are updated each month of 2020 conditionally on the new data releases which occur each month, but the parameters are not re-estimated. In total we compare the real-time performance of 12 nowcasting approaches over the pandemic months. We find that our original model with Normal homoskedastic errors produces point nowcasts as good or better than any of the other approaches. A property of Normal homoskedastic models that is often considered bad (i.e. that they are not robust to outliers), actually benefits the KMMP model as it reacts confidently to the rapidly evolving economic data. In terms of nowcast densities, we find many of the extensions lead to larger predictive variances reflecting the great uncertainty of the pandemic months. |
| Keywords: | Pandemic, Nowcasting, Income, Expenditure, Mixed frequency model, Vector Autoregression, Bayesian |
| Date: | 2021–01 |
| URL: | http://d.repec.org/n?u=RePEc:een:camaaa:2021-14&r=all |
| By: | Parley Ruogu Yang |
| Abstract: | This paper proposes a forecast-centric adaptive learning model that engages with the past studies on the order book and high-frequency data, with applications to hypothesis testing. In line with the past literature, we produce brackets of summaries of statistics from the high-frequency bid and ask data in the CSI 300 Index Futures market and aim to forecast the one-step-ahead prices. Traditional time series issues, e.g. ARIMA order selection, stationarity, together with potential financial applications are covered in the exploratory data analysis, which pave paths to the adaptive learning model. By designing and running the learning model, we found it to perform well compared to the top fixed models, and some could improve the forecasting accuracy by being more stable and resilient to non-stationarity. Applications to hypothesis testing are shown with a rolling window, and further potential applications to finance and statistics are outlined. |
| Date: | 2021–02 |
| URL: | http://d.repec.org/n?u=RePEc:arx:papers:2103.00264&r=all |
| By: | Claire Giordano (Bank of Italy); Marco Marinucci (Bank of Italy); Andrea Silvestrini (Bank of Italy) |
| Abstract: | While there is a vast macroeconomic literature that singles out the main drivers of capital accumulation in advanced economies during and after the global financial and sovereign debt crises' recessionary phase, there is much less research seeking to identify both models and variables that possess out-of-sample forecasting ability for gross fixed capital formation. Moreover, micro-founded variables are scarcely employed in macroeconomic forecasting of real investment. We fill this gap by considering a battery of univariate and multivariate time-series models to forecast investment of non-financial corporations in Italy, an interesting case-study due to its steep downturn during the two afore-mentioned crises. We find that a vector error correction model augmented with firm survey-based variables accounting for business confidence, demand uncertainty and financing constraints generally outperforms the autoregressive benchmark and a series of competing multivariate time-series models in various, alternative, evaluation samples that take into account the impact of both the global financial crisis and the sovereign debt crisis on forecast accuracy. |
| Keywords: | Real investment, forecasting evaluation, firm survey data, vector error correction model |
| JEL: | C32 C52 E22 E27 |
| Date: | 2021–02 |
| URL: | http://d.repec.org/n?u=RePEc:bdi:opques:qef_596_21&r=all |
| By: | Martin Feldkircher; Florian Huber; Gary Koop; Michael Pfarrhofer |
| Abstract: | The Panel Vector Autoregressive (PVAR) model is a popular tool for macroeconomic forecasting and structural analysis in multi-country applications since it allows for spillovers between countries in a very flexible fashion. However, this flexibility means that the number of parameters to be estimated can be enormous leading to over-parameterization concerns. Bayesian global-local shrinkage priors, such as the Horseshoe prior used in this paper, can overcome these concerns, but they require the use of Markov Chain Monte Carlo (MCMC) methods rendering them computationally infeasible in high dimensions. In this paper, we develop computationally efficient Bayesian methods for estimating PVARs using an integrated rotated Gaussian approximation (IRGA). This exploits the fact that whereas own country information is often important in PVARs, information on other countries is often unimportant. Using an IRGA, we split the the posterior into two parts: one involving own country coefficients, the other involving other country coefficients. Fast methods such as approximate message passing or variational Bayes can be used on the latter and, conditional on these, the former are estimated with precision using MCMC methods. In a forecasting exercise involving PVARs with up to $18$ variables for each of $38$ countries, we demonstrate that our methods produce good forecasts quickly. |
| Date: | 2021–03 |
| URL: | http://d.repec.org/n?u=RePEc:arx:papers:2103.04944&r=all |