The term “technical analysis” generally contains a large variety of trading techniques which are based on past movements of the asset price and a few other related variables. The use of trading rules to detect patterns in the time series of asset prices dates back to the 1800s, when traders were clearly not able to develop a fundamental analysis on the basis of extensive financial information. Persistent shifts in supply and demand had to be detected in past price movements using simple to quite elaborate techniques. Many of these techniques are still applied by practitioners, as documented in Murphy (1999).
However, the attitude of academia towards technical analysis is at least reserved, which is due to the economists' persuasion that financial markets in particular are well described by the efficient market hypothesis. Under these circumstances it is obvious that trading rules, generally not derived from a mathematically well-defined econometric or economic model, are bound not be very informative. Thus, the information content of trading signals concerning asset price fundamentals is still supposed to be largely worthless and often referred to as noise.
This view was seriously challenged by a large body of empirical studies showing that, on the one hand, standard martingale models do not sufficiently describe short-run price movements (Lewis, 1995). On the other hand, the introduction of technical analysis is justified by the results of micro survey data (Taylor and Allen, 1992), sustainable ex-post profits1, and the overall ability of heterogeneous agent models to explain the stylized facts of financial time series (Hommes, 2001). Existing explanations of the presence of chartists are generally based on sequential trading and asymmetric information2.
If news is not commonly available on financial markets, uninformed traders may infer a signal from analyzing buy and sell decisions of informed traders or changes of the asset price itself. The resulting equilibrium can be described by models of herding behavior, as is done in Banerjee (1992) and Kirman (1993). Within such asymmetric information frameworks, technical trading might also be a suitable device for informed traders. Suppose that a trader receives what he believes to be private information but he cannot be sure if the information has already been incorporated into the asset price. Before changing positions, the trader applies technical analysis to check whether his information is indeed non-public (Treynor and Ferguson, 1985).
As a general result of the models developed so far, it appears that the application of these techniques is rational from an individual trader's point of view but leads to market inefficiencies such as misalignments and excess volatility. This is due to the self-fulfilling nature of technical trading, whether or not a given initial signal is useful to predict future asset prices. However, the existing literature has not yet explicitly addressed the question as to how technical analysis might infer information about the fundamental value of the asset price as well.
To provide an information revealing explanation of technical analysis, it is assumed that information on at least some of the asset price fundamentals is available only with a considerable lag. We will argue that if the market price was indeed driven by a fundamental that is not yet observable, useful information about a possible regime shift in the stochastic process of this variable can be inferred by analyzing the asset prices themselves3.
It is shown that within such a realistic informational set up, the oscillator model described as “Hold a long position when the difference between the short term and the long term average is positive, otherwise hold a short position” (Schulmeister, 1987) carries useful information for predicting future exchange rate changes. The filter rules should be interpreted as a cheap proxy for Bayesian learning and cannot be deemed as irrational. Empirical support for this interpretation is provided by applying a Markov regime switching model to various daily US-dollar exchange rates.
The rest of the paper is organized as follows. Section 2 outlines the informational setup using a standard learning model of foreign exchange introduced by Lewis (1989). The oscillator model is derived from rational sign prediction in section 3. Section 4 reports on and discusses the estimation results and test statistics. The most important conclusions are summarized in section 5.
1 In particular, the profitability of moving average rules is repeatedly reported in the literature. For exchange rates this has been done recently by Levich and Thomas (1993), LeBaron (1999), for stocks see Brock et al. (1992) and Jegadeesh and Titman (1993, 2001).
2 Exceptions are support and resistance levels, which may have predictive power because traders prefer round numbers for stop-loss and take-profit orders (Osler, 2003).
3 Of course, there is also a related literature that derives estimates of unobservable components of an economic variable such as Friedman's 'permanent income' by means of recent observations of the variable itself (see Muth (1960) or, more recently, Barsky and De Long (1993)). Thus, the derived estimates share some properties with our solution, but the relation to technical trading has not yet been addressed.
Prof. Stefan Reitz
Next: The basic learning model
Summary: Index