| Summary: | This thesis examines the economic evaluation of forecasting strategies based on past prices, bringing together academics and practitioners techniques Forecasting methods based on past prices are convex and path-dependent dynamic strategies Therefore, they must be able to profitably exploit positive serial dependences in financial prices The most important measure of financial forecasting ability is the rate of return achieved by the predictor The expected return of forecasting strategies is first investigated by applying stochastic modelling Then, the presence of serial dependences in financial prices is tested by comparing expected and observed rates of returns of forecasting strategies According to the academic literature, the expected return of investment strategies is best established by applying stochastic modelling That is done analytically for linear forecasters, assuming that the underlying process of asset returns is not only a random walk with drift but any Gaussian processes The rate of return from financial strategies is zero under the assumption of a random walk without drift, and non-zero in all the other cases Then, it is shown that many forecasting techniques used by market participants are in fact linear forecasters and consequently fall in the scope of this study. Minimising the mean squared error is a sufficient but not necessary condition to maximise returns Under the random walk without dnft assumption, error measures and profits arenegatively correlated but very few in absolute value Only the directional accuracy exhibits high degree of linear association with profits When the true Gaussian process is not known, there are cases for which a decrease in mean squared error does not imply an increase in returns Therefore the mean squared error criterion is of poor use to maximise returns when the true model is not known The directional accuracy is of no further help Market timing ability tests based on the percentage of correct forecasts have very low power in presence of low positive autocorrelations. It is why a test of the random walk hypothesis based on the joint profitability of trading rules is investigated It happens to be powerful against a broad range of linear alternatives Its ruee feature is to exhibit a power almost equal to the best of its components unknown when the true model is unknown It constitutes as well a tool to separate mean from variance non-hnear models Simple tests of adequacy of Gaussian processes are subsequently proposed from the joint profitability of trading rules Applying previous tests, the random walk hypothesis is rejected for daily exchange rates against Dollar, over the period 1982-1992 The hypothesis of normal underlying returns is very weak compared to the independence assumption Among a few Gaussian processes, the price-trend model along with some technical models appear to be the best alternatives to explain observed trading rule returns Statistical forecasters based either on ARMA(1,1) or fractional Gaussian processes do not outperform simple technical rules Taking Into account transaction costs reduce profits to zero for individual but not for institutional Investors who might have to act on strategies that assume the foreign exchange markets exhibit positive dependencies, if not inefficiencies.
|
|---|
