Forecasting Australia's real house price index: A comparison of time series and machine learning methods

We employ 47 different algorithms to forecast Australian log real house prices and growth rates, and compare their ability to produce accurate out-of-sample predictions. The algorithms, which are specified in both single- and multi-equation frameworks, consist of traditional time series models, machine learning (ML) procedures, and deep learning neural networks. A method is adopted to compute iterated multistep forecasts from nonlinear ML specifications. While the rankings of forecast accuracy depend on the length of the forecast horizon, as well as on the choice of the dependent variable (log price or growth rate), a few generalizations can be made. For one- and two-quarter-ahead forecasts we find a large number of algorithms that outperform the random walk with drift benchmark. We also report several such outperformances at longer horizons of four and eight quarters, although these are not statistically significant at any conventional level. Six of the eight top forecasts (4 horizons × 2 dependent variables) are generated by the same algorithm, namely a linear support vector regressor (SVR). The other two highest ranked forecasts are produced as simple mean forecast combinations. Linear autoregressive moving average and vector autoregression models produce accurate olne-quarter-ahead predictions, while forecasts generated by deep learning nets rank well across medium and long forecast horizons.     

Combining forecasts based on multiple encompassing tests in a macroeconomic core system

This paper investigates whether and to what extent multiple encompassing tests may help determine weights for forecast averaging in a standard vector autoregressive setting. To this end we consider a new test‐based procedure, which assigns non‐zero weights to candidate models that add information not covered by other models. The potential benefits of this procedure are explored in extensive Monte Carlo simulations using realistic designs that are adapted to UK and to French macroeconomic data, to which trivariate vector autoregressions (VAR) are fitted. Thus simulations rely on potential data‐generating mechanisms for macroeconomic data rather than on simple but artificial designs. We run two types of forecast ‘competitions’. In the first one, one of the model classes is the trivariate VAR, such that it contains the generating mechanism. In the second specification, none of the competing models contains the true structure. The simulation results show that the performance of test‐based averaging is comparable to uniform weighting of individual models. In one of our role model economies, test‐based averaging achieves advantages in small samples. In larger samples, pure prediction models outperform forecast averages. Copyright © 2010 John Wiley & Sons, Ltd.     

The multi‐chain Markov switching model

In many real phenomena the behaviour of a certain variable, subject to different regimes, depends on the state of other variables or the same variable observed in other subjects, so the knowledge of the state of the latter could be important to forecast the state of the former. In this paper a particular multivariate Markov switching model is developed to represent this case. The transition probabilities of this model are characterized by the dependence on the regime of the other variables. The estimation of the transition probabilities provides useful information for the researcher to forecast the regime of the variables analysed. Theoretical background and an application are shown. Copyright © 2005 John Wiley & Sons, Ltd.     

Sensitivity of univariate AR(1) time‐series forecasts near the unit root

We consider the linear time‐series model y_t=d_t+u_t(t=1,...,n), where d_t is the deterministic trend and u_t the stochastic term which follows an AR(1) process; u_t=θu_t−1+ϵ_t, with normal innovations ϵ_t. Various assumptions about the start‐up will be made. Our main interest lies in the behaviour of the l‐period‐ahead forecast y_n+1 near θ=1. Unlike in other studies of the AR(1) unit root process, we do not wish to ask the question whether θ=1 but are concerned with the behaviour of the forecast estimate near and at θ=1. For this purpose we define the sth (s=1,2) order sensitivity measure λ_l^(s) of the forecast y_n+1 near θ=1. This measures the sensitivity of the forecast at the unit root. In this study we consider two deterministic trends: d_t=β_t and d_t=β_t+β_tt. The forecast will be the Best Linear Unbiased forecast. We show that, when d_t=β_t, the number of observations has no effect on forecast sensitivity. When the deterministic trend is linear, the sensitivity is zero. We also develop a large‐sample procedure to measure the forecast sensitivity when we are uncertain whether to include the linear trend. Our analysis suggests that, depending on the initial conditions, it is better to include a linear trend for reduced sensitivity of the medium‐term forecast. Copyright © 2001 John Wiley & Sons, Ltd.     

Unemployment variation over the business cycles: a comparison of forecasting models

Asymmetry has been well documented in the business cycle literature. The asymmetric business cycle suggests that major macroeconomic series, such as a country's unemployment rate, are non‐linear and, therefore, the use of linear models to explain their behaviour and forecast their future values may not be appropriate. Many researchers have focused on providing evidence for the non‐linearity in the unemployment series. Only recently have there been some developments in applying non‐linear models to estimate and forecast unemployment rates. A major concern of non‐linear modelling is the model specification problem; it is very hard to test all possible non‐linear specifications, and to select the most appropriate specification for a particular model. Artificial neural network (ANN) models provide a solution to the difficulty of forecasting unemployment over the asymmetric business cycle. ANN models are non‐linear, do not rely upon the classical regression assumptions, are capable of learning the structure of all kinds of patterns in a data set with a specified degree of accuracy, and can then use this structure to forecast future values of the data. In this paper, we apply two ANN models, a back‐propagation model and a generalized regression neural network model to estimate and forecast post‐war aggregate unemployment rates in the USA, Canada, UK, France and Japan. We compare the out‐of‐sample forecast results obtained by the ANN models with those obtained by several linear and non‐linear times series models currently used in the literature. It is shown that the artificial neural network models are able to forecast the unemployment series as well as, and in some cases better than, the other univariate econometrics time series models in our test. Copyright © 2004 John Wiley & Sons, Ltd.     

When are Direct Multi‐step and Iterative Forecasts Identical?

Although both direct multi‐step‐ahead forecasting and iterated one‐step‐ahead forecasting are two popular methods for predicting future values of a time series, it is not clear that the direct method is superior in practice, even though from a theoretical perspective it has lower mean squared error (MSE). A given model can be fitted according to either a multi‐step or a one‐step forecast error criterion, and we show here that discrepancies in performance between direct and iterative forecasting arise chiefly from the method of fitting, and is dictated by the nuances of the model's misspecification. We derive new formulas for quantifying iterative forecast MSE, and present a new approach for assessing asymptotic forecast MSE. Finally, the direct and iterative methods are compared on a retail series, which illustrates the strengths and weaknesses of each approach. Copyright © 2015 John Wiley & Sons, Ltd.     

Are Analysts' Loss Functions Asymmetric?

Despite displaying a statistically significant optimism bias, analysts' earnings forecasts are an important input to investors’ valuation models. Understanding the possible reasons for any bias is important if information is to be extracted from earnings forecasts and used optimally by investors. Extant research into the shape of analysts' loss functions explains optimism bias as resulting from analysts minimizing the mean absolute forecast error under symmetric, linear loss functions. When the distribution of earnings outcomes is skewed, optimalforecasts can appear biased. In contrast, research into analysts' economic incentives suggests that positive and negative earnings forecast errors made by analysts are not penalized or rewarded symmetrically, suggesting that asymmetric loss functions are an appropriate characterization. To reconcile these findings, we exploit results from economic theory relating to the Linex loss function to discriminate between the symmetric linear loss and the asymmetric loss explanations of analyst forecast bias. Under asymmetric loss functions optimal forecasts will appear biased even if earnings outcomes are symmetric. Our empirical results support the asymmetric loss function explanation. Further analysis also reveals that forecast bias varies systematically across firm characteristics that capture systematic variation in the earnings forecast error distribution. Copyright © 2011 John Wiley & Sons, Ltd.     

Selection of the relevant information set for predictive relationships analysis between time series

In time series analysis, a vector Y is often called causal for another vector X if the former helps to improve the k‐step‐ahead forecast of the latter. If this holds for k=1, vector Y is commonly called Granger‐causal for X . It has been shown in several studies that the finding of causality between two (vectors of) variables is not robust to changes of the information set. In this paper, using the concept of Hilbert spaces, we derive a condition under which the predictive relationships between two vectors are invariant to the selection of a bivariate or trivariate framework. In more detail, we provide a condition under which the finding of causality (improved predictability at forecast horizon 1) respectively non‐causality of Y for X is unaffected if the information set is either enlarged or reduced by the information in a third vector Z . This result has a practical usefulness since it provides a guidance to validate the choice of the bivariate system { X , Y } in place of { X , Y , Z }. In fact, to test the ‘goodness’ of { X , Y } we should test whether Z Granger cause X not requiring the joint analysis of all variables in { X , Y , Z }. Copyright © 2002 John Wiley & Sons, Ltd.     

Forecast covariances in the linear multiregression dynamic model

The linear multiregression dynamic model (LMDM) is a Bayesian dynamic model which preserves any conditional independence and causal structure across a multivariate time series. The conditional independence structure is used to model the multivariate series by separate (conditional) univariate dynamic linear models, where each series has contemporaneous variables as regressors in its model. Calculating the forecast covariance matrix (which is required for calculating forecast variances in the LMDM) is not always straightforward in its current formulation. In this paper we introduce a simple algebraic form for calculating LMDM forecast covariances. Calculation of the covariance between model regression components can also be useful and we shall present a simple algebraic method for calculating these component covariances. In the LMDM formulation, certain pairs of series are constrained to have zero forecast covariance. We shall also introduce a possible method to relax this restriction. Copyright © 2008 John Wiley & Sons, Ltd.     

Factor forecasts for the UK

Data are now readily available for a very large number of macroeconomic variables that are potentially useful when forecasting. We argue that recent developments in the theory of dynamic factor models enable such large data sets to be summarized by relatively few estimated factors, which can then be used to improve forecast accuracy. In this paper we construct a large macroeconomic data set for the UK, with about 80 variables, model it using a dynamic factor model, and compare the resulting forecasts with those from a set of standard time‐series models. We find that just six factors are sufficient to explain 50% of the variability of all the variables in the data set. These factors, which can be shown to be related to key variables in the economy, and their use leads to considerable improvements upon standard time‐series benchmarks in terms of forecasting performance. Copyright © 2005 John Wiley & Sons, Ltd.     

支持向量机在围岩变形监测中的应用

根据支特向量机优越的非线性拟合性能，建立变形量的时间序列预测模型，滚动预测围岩变形量，提高了预测模型的训练速度和预测推广能力。该方法用于西乡-固戍盾构段围岩变形预测，并与BP神经网络预测进行比较。结果表明这种模型可预测区间较长且具有较高的准确度，能够科学地指导现场施工和监测。     

Forecasting contemporaneous aggregates and the combination of forecasts: The case of the U.K. monetary aggregates

A number of papers in recent years have investigated the problems of forecasting contemporaneously aggregated time series and of combining alternative forecasts of a time series. This paper considers the integration of both approaches within the example of assessing the forecasting performance of models for two of the U.K. monetary aggregates, £M3 and MO. It is found that forecasts from a time series model for aggregate £M3 are superior to aggregated forecasts from individual models fitted to either the components or counterparts of £M3 and that an even better forecast is obtained by forming a linear combination of the three alternatives. For MO, however, aggregated forecasts from its components prove superior to either the forecast from the aggregate itself or from a linear combination of the two.     

Modeling of frequency containment reserve prices with econometrics and artificial intelligence

The forecasting of prices for electricity balancing reserve power can essentially improve the trading positions of market participants in competitive auctions. Having identified a lack of literature related to forecasting balancing reserve prices, we deploy approaches originating from econometrics and artificial intelligence and set up a forecasting framework based on autoregressive and exogenous factors. We use SARIMAX models as well as neural networks with different structures and forecast based on a rolling one-step forecast with reestimation of the models. It turns out that the naive forecast performs reasonably well but is outperformed by the more advanced models. In addition, neural network approaches outperform the econometric approach in terms of forecast quality, whereas for the further use of the generated models the econometric approach has advantages in terms of explaining price drivers. For the present application, more advanced configurations of the neural networks are not able to further improve the forecasting performance.     

A hybrid approach with step-size aggregation to forecasting hierarchical time series

Hierarchical time series arise in various fields such as manufacturing and services when the products or services can be hierarchically structured. “Top-down” and “bottom-up” forecasting approaches are often used for forecasting such hierarchical time series. In this paper, we develop a new hybrid approach (HA) with step-size aggregation for hierarchical time series forecasting. The new approach is a weighted average of the two classical approaches with the weights being optimally chosen for all the series at each level of the hierarchy to minimize the variance of the forecast errors. The independent selection of weights for all the series at each level of the hierarchy makes the HA inconsistent while aggregating suitably across the hierarchy. To address this issue, we introduce a step-size aggregate factor that represents the relationship between forecasts of the two consecutive levels of the hierarchy. The key advantage of the proposed HA is that it captures the structure of the hierarchy inherently due to the combination of the hierarchical approaches instead of independent forecasts of all the series at each level of the hierarchy. We demonstrate the performance of the new approach by applying it to the monthly data of ‘Industrial’ category of M3-Competition as well as on Pakistan energy consumption data.     

On estimating contemporaneous quarterly regional GDP

Subnational regional jurisdictions rarely have at their disposal a reasonable array of timely statistics to monitor their economic condition. In light of this, we develop a procedure that simultaneously estimates a quarterly time series for all regions of a country based upon quarterly national and annual regional data. While other such techniques exist, we suggest a temporal error structure that eliminates possible spurious jumps. Using our approach, regional analysts should now be able to distribute national growth among regions as soon as quarterly national figures are released. In a Spanish application, we detail some practicalities of the process and show that our proposal produces better estimates than the uniregional methods often used. Copyright © 2007 John Wiley & Sons. Ltd.     

Model‐based forecast adjustment: With an illustration to inflation

This paper introduces the idea of adjusting forecasts from a linear time series model where the adjustment relies on the assumption that this linear model is an approximation of a nonlinear time series model. This way of creating forecasts could be convenient when inference for a nonlinear model is impossible, complicated or unreliable in small samples. The size of the forecast adjustment can be based on the estimation results for the linear model and on other data properties such as the first few moments or autocorrelations. An illustration is given for a first‐order diagonal bilinear time series model, which in certain properties can be approximated by a linear ARMA(1, 1) model. For this case, the forecast adjustment is easy to derive, which is convenient as the particular bilinear model is indeed cumbersome to analyze in practice. An application to a range of inflation series for low‐income countries shows that such adjustment can lead to some improved forecasts, although the gain is small for this particular bilinear time series model.     

Monthly Beta Forecasting with Low‐, Medium‐ and High‐Frequency Stock Returns

This paper evaluates the accuracy of 1‐month‐ahead systematic (beta) risk forecasts in three return measurement settings; monthly, daily and 30 minutes. It was found that the popular Fama–MacBeth beta from 5 years of monthly returns generates the most accurate beta forecast among estimators based on monthly returns. A realized beta estimator from daily returns over the prior year generates the most accurate beta forecast among estimators based on daily returns. A realized beta estimator from 30‐minute returns over the prior 2 months generates the most accurate beta forecast among estimators based on 30‐minute returns. In environments where low‐, medium‐ and high‐frequency returns are accurately available, beta forecasting with low‐frequency returns are the least accurate and beta forecasting with high‐frequency returns are the most accurate. The improvements in precision of the beta forecasts are demonstrated in portfolio optimization for a targeted beta exposure. Copyright © 2016 John Wiley & Sons, Ltd.     

Forecasting Errors,Directional Accuracy and Profitability of Currency Trading: The Case of EUR/USD Exchange Rate

We provide a comprehensive study of out‐of‐sample forecasts for the EUR/USD exchange rate based on multivariate macroeconomic models and forecast combinations. We use profit maximization measures based on directional accuracy and trading strategies in addition to standard loss minimization measures. When comparing predictive accuracy and profit measures, data snooping bias free tests are used. The results indicate that forecast combinations, in particular those based on principal components of forecasts, help to improve over benchmark trading strategies, although the excess return per unit of deviation is limited. Copyright © 2016 John Wiley & Sons, Ltd.     

Forecasting Temperature Indices Density with Time‐Varying Long‐Memory Models

The hedging of weather risks has become extremely relevant in recent years, promoting the diffusion of weather‐derivative contracts. The pricing of such contracts requires the development of appropriate models for the prediction of the underlying weather variables. Within this framework, a commonly used specification is the ARFIMA‐GARCH. We provide a generalization of such a model, introducing time‐varying memory coefficients. Our model satisfies the empirical evidence of the changing memory level observed in average temperature series, and provides useful improvements in the forecasting, simulation, and pricing issues related to weather derivatives. We present an application related to the forecast and simulation of a temperature index density, which is then used for the pricing of weather options. Copyright © 2011 John Wiley & Sons, Ltd.     

Forecast Combinations in a DSGE‐VAR Lab

We explore the benefits of forecast combinations based on forecast‐encompassing tests compared to simple averages and to Bates–Granger combinations. We also consider a new combination algorithm that fuses test‐based and Bates–Granger weighting. For a realistic simulation design, we generate multivariate time series samples from a macroeconomic DSGE‐VAR (dynamic stochastic general equilibrium–vector autoregressive) model. Results generally support Bates–Granger over uniform weighting, whereas benefits of test‐based weights depend on the sample size and on the prediction horizon. In a corresponding application to real‐world data, simple averaging performs best. Uniform averages may be the weighting scheme that is most robust to empirically observed irregularities. Copyright © 2016 John Wiley & Sons, Ltd.