STAR and ANN models: forecasting performance on the Spanish “Ibex-35” stock index

This paper studies whether it is possible to exploit the nonlinear behaviour of daily returns on the Spanish Ibex-35 stock index returns to improve forecasts over short and long horizons. In this sense, we examine the out-of-sample forecast performance of smooth transition autoregression (STAR) models and artificial neural networks (ANNs). We use one-step (obtained by using recursive and nonrecursive regressions) and multi-step-ahead forecasting methods. The forecasts are evaluated with statistical and economic criteria. In terms of statistical criteria, we compared the out-of-sample forecasts using goodness of forecast measures and various testing approaches. The results indicate that ANNs consistently surpass the random walk model and, although the evidence for this is weaker, provide better forecasts than the linear AR model and the STAR models for some forecast horizons and forecasting methods. In terms of the economic criteria, we assess the relative forecast performance in a simple trading strategy including the impact of transaction costs on trading strategy profits. The results indicate a better fit for ANN models, in terms of the mean net return and Sharpe risk-adjusted ratio, by using one-step-ahead forecasts. These results show there is a good chance of obtaining a more accurate fit and forecast of the daily stock index returns by using one-step-ahead predictors and nonlinear models, but that these are inherently complex and present a difficult economic interpretation.     

Forecasting and trading high frequency volatility on large indices

The present paper analyses the forecastability and tradability of volatility on the large S&P500 index and the liquid SPY ETF, VIX index and VXX ETN. Even though there is already a huge array of literature on forecasting high frequency volatility, most publications only evaluate the forecast in terms of statistical errors. In practice, this kind of analysis is only a minor indication of the actual economic significance of the forecast that has been developed. For this reason, in our approach, we also include a test of our forecast through trading an appropriate volatility derivative. As a method we use parametric and artificial intelligence models. We also combine these models in order to achieve a hybrid forecast. We report that the results of all three model types are of similar quality. However, we observe that artificial intelligence models are able to achieve these results with a shorter input time frame and the errors are uniformly lower comparing with the parametric one. Similarly, the chosen models do not appear to differ much while the analysis of trading efficiency is performed. Finally, we notice that Sharpe ratios tend to improve for longer forecast horizons.     

Modelling and trading the EUR/USD exchange rate at the ECB fixing

The motivation for this paper is to investigate the use of alternative novel neural network (NN) architectures when applied to the task of forecasting and trading the euro/dollar (EUR/USD) exchange rate, using the European Central Bank (ECB) fixing series with only auto-regressive terms as inputs. This is done by benchmarking four different NN designs representing a higher-order neural network (HONN), a Psi Sigma Network and a recurrent neural network with the classic multilayer perception (MLP) and some traditional techniques, either statistical such as an auto-regressive moving average model, or technical such as a moving average convergence/divergence model, plus a naïve strategy. More specifically, the trading performance of all models is investigated in a forecast and trading simulation on the EUR/USD ECB fixing time series over the period 1999–2007 using the last one and half years for out-of-sample testing, an original feature of this paper. We use the EUR/USD daily fixing by the ECB as many financial institutions are ready to trade at this level and it is therefore possible to leave orders with a bank for business to be transacted on that basis. As it turns out, the MLP does remarkably well and outperforms all other models in a simple trading simulation exercise. However, when more sophisticated trading strategies using confirmation filters and leverage are applied, the HONN network produces better results and outperforms all other NN and traditional statistical models in terms of annualized return.     

A neural network enhanced volatility component model

Volatility prediction, a central issue in financial econometrics, attracts increasing attention in the data science literature as advances in computational methods enable us to develop models with great forecasting precision. In this paper, we draw upon both strands of the literature and develop a novel two-component volatility model. The realized volatility is decomposed by a nonparametric filter into long- and short-run components, which are modeled by an artificial neural network and an ARMA process, respectively. We use intraday data on four major exchange rates and a Chinese stock index to construct daily realized volatility and perform out-of-sample evaluation of volatility forecasts generated by our model and well-established alternatives. Empirical results show that our model outperforms alternative models across all statistical metrics and over different forecasting horizons. Furthermore, volatility forecasts from our model offer economic gain to a mean-variance utility investor with higher portfolio returns and Sharpe ratio.     

Nonlinear prediction of exchange rates with monetary fundamentals

This paper employs a neural network (NN) to study the nonlinear predictability of exchange rates for four currencies at the 1-, 6- and 12-month forecast horizons. We find that our neural network model with market fundamentals cannot beat the random walk (RW) in out-of-sample forecast accuracy, although it occasionally shows a limited market-timing ability. The neural network model without monetary fundamentals forecasts somewhat better for the British pound and the Canadian dollar. The model also exhibits some market-timing ability for the Deutsche mark at the 6- and 12-month horizons, and for the Canadian dollar at the 1-month horizon. In general, the model performs more poorly when it becomes more complex or when the forecast horizon lengthens. Our overall results are more on the negative side and suggest that neither nonlinearity nor market fundamentals appear to be very important in improving exchange rate forecast for the chosen horizons.     

Forecasting the term structure of Chinese Treasury yields

This paper is the first to study the forecasting of the term structure of Chinese Treasury yields. We extend the Nelson–Siegel class of models to estimate and forecast the term structure of Chinese Treasury yields. Our empirical analysis shows that the models fit the data very well, and that more flexible specifications dramatically improve in-sample fitting performance. In particular, the model which enhances slope fitting is the best in capturing the Chinese yield curve dynamics. We also demonstrate that time-varying factors of the models may be interpreted as the level, slope and curvature of the yield curve. Furthermore, we use five dynamic processes for the time-varying factors to forecast the term structure at both short and long horizons. Our forecasts are much more accurate than the random walk, the Cochrane–Piazzesi regression and the AR(1) benchmark models at long horizons.     

USING NEURAL NETS TO COMBINE INFORMATION SETS IN CORPORATE BANKRUPTCY PREDICTION

We demonstrate that the use of a neural network (NN) model to combine information from corporate financial statements and equity markets provides improved predictive estimates of the probability of corporate bankruptcy. Using performance measures, based on the receiver operating characteristic curve, the forecast combinations from the NN models are demonstrated to outperform the forecasts derived from a forecast combination generated using a logistic regression approach. This result provides support for the use of forecast combinations generated from NN models in the estimation of corporate bankruptcy probabilities as it outperforms the standard approach of forming a hybrid forecasting model which includes all the explanatory variables. Copyright © 2012 John Wiley & Sons, Ltd.     

Stock volatility predictability in bull and bear markets

The recent literature on stock return predictability suggests that it varies substantially across economic states, being strongest during bad economic times. In line with this evidence, we document that stock volatility predictability is also state dependent. In particular, in this paper, we use a large data set of high-frequency data on individual stocks and a few popular time-series volatility models to comprehensively examine how volatility forecastability varies across bull and bear states of the stock market. We find that the volatility forecast horizon is substantially longer when the market is in a bear state than when it is in a bull state. In addition, over all but the shortest horizons, the volatility forecast accuracy is higher when the market is in a bear state. This difference increases as the forecast horizon lengthens. Our study concludes that stock volatility predictability is strongest during bad economic times, proxied by bear market states.     

A RADIAL BASIS FUNCTION APPROACH TO EARNINGS FORECAST

The fundamental management problem of decision making in a climate where future values of important variables are unknown and can at best be estimated using traditional statistical techniques is addressed. The incorporation of forecast models into management decision‐support systems is critical for the overall success of organizational accounting information systems, where managers require confidence in the information that they use. The neural network paradigm has been described as a promising nonparametric approach, negating the required, and sometimes restrictive, statistical assumptions. The application of the neural network paradigm to the area of earnings forecasting is presented. A radial basis function (RBF) approach is developed and tested empirically using data from the Hong Kong Hang Seng 100 Index and macroeconomic data, mimicking an actual business valuation/forecast exercise. Results show that the RBF approach is superior to regression and financial analysts in earnings forecast. Copyright © 2012 John Wiley & Sons, Ltd.     

Forecasting volatility of crude oil futures using a GARCH–RNN hybrid approach

Volatility is an important element for various financial instruments owing to its ability to measure the risk and reward value of a given financial asset. Owing to its importance, forecasting volatility has become a critical task in financial forecasting. In this paper, we propose a suite of hybrid models for forecasting volatility of crude oil under different forecasting horizons. Specifically, we combine the parameters of generalized autoregressive conditional heteroscedasticity (GARCH) and Glosten–Jagannathan–Runkle (GJR)-GARCH with long short-term memory (LSTM) to create three new forecasting models named GARCH–LSTM, GJR-LSTM, and GARCH-GJRGARCH LSTM in order to forecast crude oil volatility of West Texas Intermediate on different forecasting horizons and compare their performance with the classical volatility forecasting models. Specifically, we compare the performances against existing methodologies of forecasting volatility such as GARCH and found that the proposed hybrid models improve upon the forecasting accuracy of Crude Oil: West Texas Intermediate under various forecasting horizons and perform better than GARCH and GJR-GARCH, with GG–LSTM performing the best of the three proposed models at 7-, 14-, and 21-day-ahead forecasts in terms of heteroscedasticity-adjusted mean square error and heteroscedasticity-adjusted mean absolute error, with significance testing conducted through the model confidence set showing that GG–LSTM is a strong contender for forecasting crude oil volatility under different forecasting regimes and rolling-window schemes. The contribution of the paper is that it enhances the forecasting ability of crude oil futures volatility, which is essential for trading, hedging, and purposes of arbitrage, and that the proposed model dwells upon existing literature and enhances the forecasting accuracy of crude oil volatility by fusing a neural network model with multiple econometric models.     

PROBABILISTIC APPROACHES FOR CREDIT SCREENING AND BANKRUPTCY PREDICTION

Three probabilistic neural network approaches are used for credit screening and bankruptcy prediction: a logistic regression neural network (LRNN), a probabilistic neural network (PNN) and a semi‐supervised expectation maximization‐based neural network. Using real‐world bankruptcy prediction and credit screening datasets, we compare the three probabilistic approaches using various performance criteria of sensitivity, specificity, accuracy, decile lift and area under receiver operating characteristics (ROC) curves. The results of our experiments indicate that the PNN outperforms the other two techniques for decile lift and specificity performance metric. Using the area under ROC curve, we find that for bankruptcy prediction data the PNN outperforms the other two approaches when false positive rates (FPRs) are less than 40 %. LRNN outperforms the other two techniques for FPRs higher than 40 % for bankruptcy data. We observe that the LRNN results are very sensitive to the ratio of examples belonging to two classes in training data and there is a tendency to overfit training data. Copyright © 2012 John Wiley & Sons, Ltd.     

Adaptive Methods in Macroeconomic Forecasting

Adaptive methods are used to forecast three main Austrian economic indicators. We use a weighted recursive model as well as a neural network approach both with and without adaptive characteristics and compare our results to the forecasts of two Austrian research institutes. It appears that even models which use very limited information can outperform the two Institutes’ forcasts of the unemployment rate. For the case of most series adaptivity represents a possibility of improving the forecasts. © 1997 by John Wiley & Sons, Ltd.     

Why do forecasters disagree? Lessons from the term structure of cross-sectional dispersion

Key sources of disagreement among economic forecasters are identified by using data on cross-sectional dispersion in forecasters’ long- and short-run predictions of macroeconomic variables. Dispersion among forecasters is highest at long horizons where private information is of limited value and lower at short forecast horizons. Moreover, differences in views persist through time. Such differences in opinion cannot be explained by differences in information sets; our results indicate they stem from heterogeneity in priors or models. Differences in opinion move countercyclically, with heterogeneity being strongest during recessions where forecasters appear to place greater weight on their prior beliefs.     

Using Neural Data to Test a Theory of Investor Behavior: An Application to Realization Utility

We conduct a study in which subjects trade stocks in an experimental market while we measure their brain activity using functional magnetic resonance imaging. All of the subjects trade in a suboptimal way. We use the neural data to test a “realization utility” explanation for their behavior. We find that activity in two areas of the brain that are important for economic decision‐making exhibit activity consistent with the predictions of realization utility. These results provide support for the realization utility model. More generally, they demonstrate that neural data can be helpful in testing models of investor behavior.     

Smooth transition patterns in the realized stock–bond correlation

This paper explores the time variation in the stock–bond correlation using high-frequency data. Gradual transitions between regimes of negative and positive stock–bond correlation are well accommodated by the smooth transition regression (STR) model. We find that the regimes are systematically related to movements in financial and to a minor extent macroeconomic transition variables. In particular, the most informative transition variables are the short rate, the yield spread, and the VIX volatility index. Importantly, both in-sample and out-of-sample evaluation criteria show that multiple transition variable STR specifications considerably outperform single transition variable STR models. Our results are robust to different forecast horizons.     

The Taylor Rule and Forecast Intervals for Exchange Rates

In this paper, we examine the Meese–Rogoff puzzle from a different perspective: out‐of‐sample interval forecasting. While most studies in the literature focus on point forecasts, we apply semiparametric interval forecasting to a group of exchange rate models. Forecast intervals for 10 OECD exchange rates are generated and the performance of the empirical exchange rate models are compared with the random walk. Our contribution is twofold. First, we find that in general, exchange rate models generate tighter forecast intervals than the random walk, given that their intervals cover out‐of‐sample exchange rate realizations equally well. Our results suggest a connection between exchange rates and economic fundamentals: economic variables contain information useful in forecasting distributions of exchange rates. We also find that the benchmark Taylor rule model performs better than the monetary, PPP and forward premium models, and its advantages are more pronounced at longer horizons. Second, the bootstrap inference framework proposed in this paper for forecast interval evaluation can be applied in a broader context, such as inflation forecasting.     

Moments of multivariate regime switching with application to risk-return trade-off

We use a Fourier transform to derive multivariate conditional and unconditional moments of multi-horizon returns under a regime-switching model. These moments are applied to examine the relevance of risk horizon and regimes for buy-and-hold investors. We analyze the impact of time-varying expected returns and risk (variance and covariance) on portfolio allocations' “term structure”—portfolio allocations as a function of the investment horizon. Using monthly observations on S&P composite index and 10-year Government Bond, we find that the term structure of the optimal allocations depends on market conditions measured by the probability of being in bull state. At short horizons and when this probability is low, buy-and-hold investors decrease their holdings of risky assets. We also find that the conditional optimal portfolio performs quite well at short and intermediate horizons and less at long horizons.     

Forecasting currency volatility: A comparison of implied volatilities and AR(FI)MA models

We compare forecasts of the realized volatility of the pound, mark and yen exchange rates against the dollar, calculated from intraday rates, over horizons ranging from one day to three months. Our forecasts are obtained from a short memory ARMA model, a long memory ARFIMA model, a GARCH model and option implied volatilities. We find intraday rates provide the most accurate forecasts for the one-day and one-week forecast horizons while implied volatilities are at least as accurate as the historical forecasts for the one-month and three-month horizons. The superior accuracy of the historical forecasts, relative to implied volatilities, comes from the use of high frequency returns, and not from a long memory specification. We find significant incremental information in historical forecasts, beyond the implied volatility information, for forecast horizons up to one week.     

GP algorithm versus hybrid and mixed neural networks

In the current paper, we present an integrated genetic programming (GP) environment called java GP modelling. The java GP modelling environment is an implementation of the steady-state GP algorithm. This algorithm evolves tree-based structures that represent models of inputs and outputs. The motivation of this paper is to compare the GP algorithm with neural network (NN) architectures when applied to the task of forecasting and trading the ASE 20 Greek Index (using autoregressive terms as inputs). This is done by benchmarking the forecasting performance of the GP algorithm and six different autoregressive moving average model (ARMA) NN combination designs representing a Hybrid, Mixed Higher Order Neural Network (HONN), a Hybrid, Mixed Recurrent Neural Network (RNN), a Hybrid, Mixed classic Multilayer Perceptron with some traditional techniques, either statistical such as a an ARMA or technical such as a moving average convergence/divergence model, and a naïve trading strategy. More specifically, the trading performance of all models is investigated in a forecast and trading simulation on ASE 20 time-series closing prices over the period 2001–2008, using the last one and a half years for out-of-sample testing. We use the ASE 20 daily series as many financial institutions are ready to trade at this level, and it is therefore possible to leave orders with a bank for business to be transacted on that basis. As it turns out, the GP model does remarkably well and outperforms all other models in a simple trading simulation exercise. This is also the case when more sophisticated trading strategies using confirmation filters and leverage are applied, as the GP model still produces better results and outperforms all other NN and traditional statistical models in terms of annualized return.