AUD/USD "bipolar disorder" history 
Written by Forex Automaton  
Wednesday, 29 October 2008 17:36  
This is a followup note to AUD/USD: predictability overview. AUD/USD is one of the currency pairs with what I call hourscale "bipolar disorder" predictability feature: a tendency to form quickly alternating rises and falls on nexthour time scale, more pronounced than in a fully unpredictable time series of same volatility. Any statistical test of market inefficiency requires a finite time of observation, and the original article covered the time interval from August 20, 2002 to February 1st, 2008, with no attention paid to the time evolution of the picture of predictable patterns in this currency pair. This article extends the historical coverage into the first three quarters of 2008, and focuses on the time evolution of the pattern, characterized by the 1hour time lag autocorrelation value. Naturally, we want to know how robust this feature has been, whether it is alive at present, and what are its future prospects. For the sake of completeness, Fig.1 presents the history of AUD/USD for the period under study. Fig.2 portrays the subject of this note  the "bipolar disorder", using the psychiatric analogy ascribed to Benjamin Graham ("Intelligent Investor"). Justifying the psychiatric analogy, the large negative autocorrelation signal in the bin next to zero indicates rapid (next hour) changes in the mood of the market, a price action followed by a next hour correction. Technically, these represent a profit opportunity to a speculator able to stay cool and take advantage of the market's excesses, entering after the "action" and taking profit on the predictable "reaction". What is not clear from Fig.2 is whether these events happen systematically.
To trace the time evolution of the effect, Fig.3 plots the correlation value at the 1 hour time lag over time. Plotted as red background is the noise estimate, obtained from martingale simulations based on the historical volatilities of AUD/USD for the period under study. The noise is presented as mean plusminus 1 RMS, where the RMS characterizes distribution of the correlation value obtained for this particular time lag bin by analyzing 20 independent simulated pairs of uncorrelated time series. The RMS is a measure of accuracy in the determination of the correlation values, an uncertainty dependent on the amount of data and the time scale. This shows that the feature is statistically significant.
Fig.4, time evolution of variance, demonstrates that the effect in Fig.3 is even more significant than it looks  as the next figure will show, the average volatility for the period is strongly influenced by the most recent data, 3rd quarter of 2008, and is therefore an overestimate for the overall period 20022008). There is another important message: volatility varies strongly and apparently, the time variation of the strength of the effect is entirely due to a change in the volatility with time. Checking this is best done by representing the effect not as an absolute correlation, but as a Pearson correlation coefficient  covariance normalized to the variance.
Presenting the normalized magnitude of the one hour time lag autocorrelation as a function of time, Fig.5 has a few important messages:
Comparison of the time histories of the LIBOR rate differential, Fig.6, and variance, a measure of volatility, Fig.4, reveals that the two closely follow one another  interest rate differential creates volatility. It would be harder to say the same about the timelag 1 effect  at least with the European trading data from Fig.5  the effect appears and disappears, but at least one thing is certain: there is either the negative predictive correlation or no predictive correlation  a positive correlation such as would be required to justify momentumtrading or trend following in AUD/JPY, is not there. 

Last Updated ( Monday, 14 September 2009 16:52 ) 