This thesis involved an empirical investigation of the predictability of Australian industrial stock returns using a dynamic state-space framework. The systematic risks of industrial portfolios were examined in a stochastic market- model. The systematic risks of industry portfolios are found to be stochastic processes. Most of the industry groups have time-varying systematic risks that are mean-reverting to their stable or moving long-term mean. However, the investment and financial services, alcohol and tobacco, gold, insurance and media industry groups have rather random systematic risks. The time-varying market model provides a better explanation of the portfolio returns than the single-index model since it captures the stochastic properties of market risk. Further, a Bayesian dynamic-forecasting model was employed to examine the explanatory power of a set of economic and financial variables. The unanticipated components of the term-structure variable, the interest-rate variable and the aggregate-dividend-yield variable were shown to be significant in explaining the industry portfolio excess returns. The comparison between multivariate analysis and univariate analysis strongly indicates that the correlations within industries are critical in the investigation of the predictability of returns. In the out-of-sample analysis, a maximally predicted portfolio (MPP) was constructed based on the updated economic and financial information; however, the predictability of the MPP did not exceed that of a naive forecast. / Furthermore, the market timing ability associated with the predictability of the MPP was insignificant. The industry-group-rotation strategy is able to enhance the industry portfolio performance, but the predictability only contributes a small proportion of the profits. The results indicate that the industry returns contain predictive components; however, investors are less likely to exploit the existing predictability to gain excess profit. The level of predictability discovered here does not contradict market-efficiency theory.

Le but de cette thèse est de construire un modèle de détection des changements de phase -passages de marché haussier à baissier et vice versa - du marché des actions américaines cotées, en utilisant un nombre relativement important de variables à la fois fondamentales (macroéconomiques et microéconomiques) et issues de l’analyse technique.Le modèle statistique retenu est la régression logistique statique, avec un retard pour les variables explicatives allant de zéro à trois mois. Les huit variables les plus significatives parmi vingt candidatesont été sélectionnées à partir des données mensuelles du S&P500 sur la période 1963-2003. Le modèle obtenu a été testé sur 2004-2013 et sa performance a été supérieure à celles de la stratégie Buy & Holdet d’un modèle univarié utilisant la variable ayant le plus fort pouvoir de détection - ce dernier modèle ayant fait l’objet d’une étude dans la littérature.Il a également été montré que des variables non encore considérées dans la littérature - la moyenne mobile sur les six derniers mois des créations nettes d’emplois non-agricoles, la base monétaire et le Composite Leading Indicator de l’OCDE - ont un pouvoir de détection significatif pour notre problématique. D'autre part, la variable binaire indiquant la position du S&P500 par rapport à sa moyenne mobile des dix derniers mois - variable de type analyse technique - a un pouvoir prédictif beaucoup plus élevé que les variables fondamentales étudiées. Enfin, les deux autres variables les plus statistiquement significatives sont macroéconomiques : l'écart entre les taux à dix ans des T-bonds et à trois mois des T-bills et la moyenne mobile des créations d’emplois non-agricoles. / The goal of this thesis is to build a model capable of detecting the reversals - shift from bull market to bear market or vice versa - of the American stock market, by using a relatively large number of explanatory variables, both of fundamental (macroeconomic and microeconomic) and of ‘technical analysis’ types.The statistical model used is static logistic regression, with lags for the independent variables ranging from zero to three months. Starting with twenty variables, the eight most significant ones have been selected on a training set consisting of monthly data of the S&P500 between 1963 and 2003. There sulting model has been tested over the 2004-2013 period and its performance was better than those of a buy & hold strategy and of a univariate model based on the variable with the highest predictive power – the latter model being the focus of a paper in the current literature. Another contribution of the thesis is that some variables not yet studied in the literature – the six month moving average of net non-farm job creations, the monetary base and the OECD Composite Leading Indicator – are statistically significant for our problem. Moreover, the predictive power of the binary variable indicating whether the S&P500 is above or below its ten-month moving average – a technical analysis variable – is much higher than that of the fundamental variables which have been considered. Finally, the two other most significant variables are macroeconomic ones: the spread between the ten-year T-bond and three-month T-bill rates and the moving average of non-farm jobs creations.