AUD/JPY and GBP/USD leading indicator history, 20022009 
Written by Forex Automaton  
Monday, 20 April 2009 16:41  
AUD/JPY and GBP/USD form a weakly correlated pair, often but not always correlated positively, which many traders would regard to be one of the last suspects when it comes to predictive forex correlations. Nonetheless, in the screening we conducted last summer using the data set from 2002 to early 2008, AUD/JPY was found to form a leading indicator for GBP/USD. The feature was found by studying the timeintegrated correlation of logarithmic returns in the two time series. As usual in such cases, a detailed timeevolution study is necessary to tell whether this effect is merely a result of a single highimpact event or a recurrent feature. I extend the period of observation up to April 2009, split it into three time windows of varying volatility, and analyze the time stability of the leading indicator effect.
By visual inspection of the AUD/JPY and GBP/USD charts, Fig.1, it's impossible to tell the "instantaneous" correlation, devoid of predictive value, from predictive correlations (those associated with nonzero time lags). The charts look somewhat positively correlated or in other words the two time series sometimes move in tandem. Quantitative analysis is needed to measure the magnitude of the correlation and the range of time lags affected by it.
Fig.2 presents the history of correlation between AUD/JPY and GBP/USD, including the nonzero time lags. The change in volatility can be seen by looking at the history of the 0lag correlation magnitude, a fair measure of variance in this situation (of nearzero average time series element). Correlations with nonzero time lags are of particular importance for the purpose of forex trading system development, since their presence indicates predictability of one time series on the basis on the other. The correlation value associated with 1 hour time lag is seen to be strongly biased in the positive direction. It is clear that in the original report, we were not dealing with the impact of just one or two events unlikely to reproduce themselves in the future, but rather with an effect distributed in time.
Statistical significance of the effect is analyzed below in the usual manner, and as usual, the quality of the analysis depends on how the nonstationarity of the time series is being treated. Fig.3 shows how we split the time series into the fragments A,B, and C, each being roughly stationary. Indeed, considering each piece as stationary is much better than considering the entire time series as stationary. What is conventionally regarded as the present financial crisis, began in fragment B. Fragment C corresponds to the most painful phase of the crisis experienced so far. Fig.4:Crosscorrelation of AUD/JPY and GBP/USD is shown against the backdrop of statistical noise (red). The noise is obtained from martingale simulations based on the recorded volatilities of EUR/JPY and USD/CHF in this trading session for the period under study. The noise is presented as mean plusminus 1 RMS, where RMS characterizes the distribution of the correlation value obtained for each particular bin by analyzing 20 independent simulated pairs of uncorrelated time series. A,B and C fragments are defined as shown in Fig.3. Fig.4 is our typical figure to estimate statistical significance of the nontrivial correlation. It shows the leadingindicator feature to be significant in fragments A and B. In fragment C, the "impact phase" of the crsis, it is not significant. This is not out of line with the observations of the modifications of the forex autocorrelations during the crisis, the net observation being a reduction in the trend sustainability and evolution towards "bipolar" behavior, not necessarily reaching significant bipolarity. 

Last Updated ( Monday, 04 January 2010 12:36 ) 