An ensemble of LSTM neural networks for high‐frequency stock market classification

We propose an ensemble of long–short‐term memory (LSTM) neural networks for intraday stock predictions, using a large variety of technical analysis indicators as network inputs. The proposed ensemble operates in an online way, weighting the individual models proportionally to their recent performance, which allows us to deal with possible nonstationarities in an innovative way. The performance of the models is measured by area under the curve of the receiver operating characteristic. We evaluate the predictive power of our model on several US large‐cap stocks and benchmark it against lasso and ridge logistic classifiers. The proposed model is found to perform better than the benchmark models or equally weighted ensembles.     

Dynamic Model Averaging and CPI Inflation Forecasts: A Comparison between the Euro Area and the United States

The paper forecasts consumer price inflation in the euro area (EA) and in the USA between 1980:Q1 and 2012:Q4 based on a large set of predictors, with dynamic model averaging (DMA) and dynamic model selection (DMS). DMA/DMS allows not solely for coefficients to change over time, but also for changes in the entire forecasting model over time. DMA/DMS provides on average the best inflation forecasts with regard to alternative approaches (such as the random walk). DMS outperforms DMA. These results are robust for different sample periods and for various forecast horizons. The paper highlights common features between the USA and the EA. First, two groups of predictors forecast inflation: temporary fundamentals that have a frequent impact on inflation but only for short time periods; and persistent fundamentals whose switches are less frequent over time. Second, the importance of some variables (particularly international food commodity prices, house prices and oil prices) as predictors for consumer price index inflation increases when such variables experience large shocks. The paper also shows that significant differences prevail in the forecasting models between the USA and the EA. Such differences can be explained by the structure of these respective economies. Copyright © 2015 John Wiley & Sons, Ltd.     

A ground‐level ozone forecasting model for Santiago,Chile

A physically based model for ground‐level ozone forecasting is evaluated for Santiago, Chile. The model predicts the daily peak ozone concentration, with the daily rise of air temperature as input variable; weekends and rainy days appear as interventions. This model was used to analyse historical data, using the Linear Transfer Function/Finite Impulse Response (LTF/FIR) formalism; the Simultaneous Transfer Function (STF) method was used to analyse several monitoring stations together. Model evaluation showed a good forecasting performance across stations—for low and high ozone impacts—with power of detection (POD) values between 70 and 100%, Heidke's Skill Scores between 40% and 70% and low false alarm rates (FAR). The model consistently outperforms a pure persistence forecast. Model performance was not sensitive to different implementation options. The model performance degrades for two‐ and three‐days ahead forecast, but is still acceptable for the purpose of developing an environmental warning system at Santiago. Copyright © 2002 John Wiley & Sons, Ltd.     

Time‐simultaneous prediction band for a time series

I propose principles and methods for the construction of a time‐simultaneous prediction band for a univariate time series. The methods are entirely based on a learning sample of time trajectories, and make no parametric assumption about its distribution. Hence, the methods are general and widely applicable. The expected coverage probability of a band can be estimated by a bootstrap procedure. The estimate is likely to be less than the nominal level. Expected lack of coverage can be compensated for by increasing the coverage in the learning sample. Applications to simulated and empirical data illustrate the methods. Copyright © 2007 John Wiley & Sons, Ltd.     

Forecasting volatility by means of threshold models

The aim of this paper is to compare the forecasting performance of competing threshold models, in order to capture the asymmetric effect in the volatility. We focus on examining the relative out‐of‐sample forecasting ability of the SETAR‐Threshold GARCH (SETAR‐TGARCH) and the SETAR‐Threshold Stochastic Volatility (SETAR‐THSV) models compared to the GARCH model and Stochastic Volatility (SV) model. However, the main problem in evaluating the predictive ability of volatility models is that the ‘true’ underlying volatility process is not observable and thus a proxy must be defined for the unobservable volatility. For the class of nonlinear state space models (SETAR‐THSV and SV), a modified version of the SIR algorithm has been used to estimate the unknown parameters. The forecasting performance of competing models has been compared for two return time series: IBEX 35 and S&P 500. We explore whether the increase in the complexity of the model implies that its forecasting ability improves. Copyright © 2007 John Wiley & Sons, Ltd.     

Unit roots versus other types of time heterogeneity,parameter time dependence and superexogeneity

This paper stresses the restrictive nature of the standard unit root/cointegration assumptions and examines a more general type of time heterogeneity, which might characterize a number of economic variables, and which results in parameter time dependence and misleading statistical inference. We show that in such cases ‘operational’ models cannot be obtained, and the estimation of time‐varying parameter models becomes necessary. For instance, economic processes subject to endemic change can only be adequately modelled in a state space form. This is a very important point, because unstable models will break down when used for forecasting purposes. We also discuss a new test for the null of cointegration developed by Quintos and Phillips (1993), which is based on parameter constancy in cointegrating regressions. Finally, we point out that, if it is possible to condition on a subset of superexogenous variables, parameter instability can be handled by estimating a restricted system. Copyright © 2002 John Wiley & Sons, Ltd.     

The performance of non‐linear exchange rate models: a forecasting comparison

In recent years there has been a considerable development in modelling non‐linearities and asymmetries in economic and financial variables. The aim of the current paper is to compare the forecasting performance of different models for the returns of three of the most traded exchange rates in terms of the US dollar, namely the French franc (FF/$), the German mark (DM/$) and the Japanese yen (Y/$). The relative performance of non‐linear models of the SETAR, STAR and GARCH types is contrasted with their linear counterparts. The results show that if attention is restricted to mean square forecast errors, the performance of the models, when distinguishable, tends to favour the linear models. The forecast performance of the models is evaluated also conditional on the regime at the forecast origin and on density forecasts. This analysis produces more evidence of forecasting gains from non‐linear models. Copyright © 2002 John Wiley & Sons, Ltd.     

Forecasting the Daily Time‐Varying Beta of European Banks During the Crisis Period: Comparison Between GARCH Models and the Kalman Filter

This intention of this paper is to empirically forecast the daily betas of a few European banks by means of four generalized autoregressive conditional heteroscedasticity (GARCH) models and the Kalman filter method during the pre‐global financial crisis period and the crisis period. The four GARCH models employed are BEKK GARCH, DCC GARCH, DCC‐MIDAS GARCH and Gaussian‐copula GARCH. The data consist of daily stock prices from 2001 to 2013 from two large banks each from Austria, Belgium, Greece, Holland, Ireland, Italy, Portugal and Spain. We apply the rolling forecasting method and the model confidence sets (MCS) to compare the daily forecasting ability of the five models during one month of the pre‐crisis (January 2007) and the crisis (January 2013) periods. Based on the MCS results, the BEKK proves the best model in the January 2007 period, and the Kalman filter overly outperforms the other models during the January 2013 period. Results have implications regarding the choice of model during different periods by practitioners and academics. Copyright © 2016 John Wiley & Sons, Ltd.     

Forecasting VaR models under Different Volatility Processes and Distributions of Return Innovations

This paper provides clear‐cut evidence that the out‐of‐sample VaR (value‐at‐risk) forecasting performance of alternative parametric volatility models, like EGARCH (exponential general autoregressive conditional heteroskedasticity) or GARCH, and Markov regime‐switching models, can be considerably improved if they are combined with skewed distributions of asset return innovations. The performance of these models is found to be similar to that of the EVT (extreme value theory) approach. The performance of the latter approach can also be improved if asset return innovations are assumed to be skewed distributed. The performance of the Markov regime‐switching model is considerably improved if this model allows for EGARCH effects, for all different volatility regimes considered. Copyright © 2014 John Wiley & Sons, Ltd.     

A Simple Linear Regression Approach to Modeling and Forecasting Mortality Rates

Observing that a sequence of negative logarithms of 1‐year survival probabilities displays a linear relationship with the sequence of corresponding terms with a time lag of a certain number of years, we propose a simple linear regression to model and forecast mortality rates. Our model assuming the linearity between two mortality sequences with a time lag each other does not need to formulate the time trends of mortality rates across ages for mortality prediction. Moreover, the parameters of our model for a given age depend on the mortality rates for that age only. Therefore, whether the span of the study ages with the age included is widened or shortened will not affect the results of mortality fitting and forecasting for that age. In the empirical testing, the regression results using the mortality data for the UK, USA and Japan show a satisfactory goodness of fit, which convinces us of the appropriateness of the linear assumption. Empirical illustrations further show that our model's performances of fitting and forecasting mortality rates are quite satisfactory compared with the existing well‐known mortality models. Copyright © 2015 John Wiley & Sons, Ltd.