Exchange rates expectations and chaotic dynamics: a replication study

In this paper the author analyzes the behavior of exchange rates expectations for four currencies, by considering a re-calculation and an extension of Resende and Zeidan (Expectations and chaotic dynamics: empirical evidence on exchange rates, Economics Letters, 2008). Considering Lyapunov exponent-based tests results, they are not supportive of chaos in exchange rates expectations, although the so-called 0−1 test strongly supports the chaos hypothesis. (Published in Special Issue Replication Study) JEL C12 C15

1 Introduction Resende and Zeidan (2008) test for deterministic chaos in exchange rates expectations for four currencies, by means of Fernández-Rodríguez et al. (2005) test, F SA hereafer. Their work, however, presents some formal drawbacks. In the next sections, We re-analyze the problem at hand more accurately and we apply alternative statistical techniques.
Methods are described on sections 2 through 4. Section 5 briefly reviews the original paper and section 6 presents the results with a discussion. Section 7 concludes.

F SA test
The null hypothesis of the test is deterministic chaos. Given a time series of length T , {y t } T t=1 , F SA suggest the estimation of the dominant Lyapunov exponent, λ, to be the bootstrap aggregation, or bagging, of 100 block bootstrap samples of the largest Lyapunov exponents.
Since the largest Lyapunov exponent does not increase when increasing the sample size in a deterministic process, but increases otherwise, F SA propose using λ max (T i ) to test for the stability of the largest Lyapunov exponent, where T i is the size of block i. This non-increasing property of the largest Lyapunov exponent with the sample size for chaotic processes may be tested by the regression: for T i = T 1 , . . . , T r = T , where r is the number of blocks. Now our interest is on testing the null hypothesis H 0 : α ≤ 0. The statistic isα 1 /s 1 where s 1 is the standard error ofα 1 , computed through the Newey-West HAC covariance matrix estimator. 1 In order to estimate the Lyapunov exponents, we will compute Rosenstein's et al. (1995) algorithm with the parameters selected as suggested in F SA (2005).

Shintani and Linton (2004) test
Let {y t } T t=1 , be a random scalar sequence generated by the nonlinear autoregressive model: where θ 0 : R d → R is a nonlinear dynamic map and {u t } is a sequence of random variables. The model can be expressed in terms of a map with an error vector U t = (u t , 0, . . . , 0) and the map function F : Let J t be the Jacobian of the map F evaluated at Z t . We can obtain the neural network estimator of the largest Lyapunov exponent. Now we distinguish between the sample size T used for estimating the JacobianĴ t and the block length M , which is the number of evaluation points used for estimating the Lyapunov exponent.
Our interest is to test the null hypothesis H 0 : λ ≥ 0 against the alternative H 1 : λ < 0, therefore the test is one-sided. Given a consistent estimate tandard deviation of the estimatorλ,Φ, the test statistic is: Under the null hypothesis,t is asymptotically distributed as a N (0, 1). We compute the heteroskedasticity and autocorrelation consistent (HAC) covariance matrix estimator. 2 Three alternative selection procedures can be applied: Full, Block and ES (equally spaced subsamples, see Shintani and Linton, 2004, for the details). In all cases, given the number of hidden units (r), the lag length (d) has been selected based on the Bayesian information criterion. The quadratic spectral kernel with optimal bandwidth has been used for the heteroskedasticity and autocorrelation consistent covariance estimation.
2. Compute the mean square displacement of the transition variables for several values of c ∈ (0, π) and for n = 1, 2, . . . , n cut , where n cut N : 3. Form the vectors ξ = (1, 2, . . . n cut ) and ∆ = (D c (1), D c (2), . . . , D c (n cut )) and estimate the asymptotic growth rate as the sample correlation coefficient: Webel (2012) computed the test on DAX asset returns, assuming that asset returns time series {r t } N t=1 are defined by: where {s t } is the one-dimensional signal, an observable of an underlying k−dimensional deterministic dynamical system, and ε t is the random noise. Webel suggests filtering out noise by wavelet denoising techniques, and then applying the 0-1 test to the denoised series. More specifically, Webel (2012) computed Maximal Overlap Discrete Wavelet Transforms (MODWT) for each returns series of 1655 observations, with four Daubechies filters. In this work we follow his guidelines, and compute the denoised original RZ residuals for two filters: Coiflet and Haar and a coarsest resolution chosen to be [log 2 (167)] = 7 and a soft thresholding rule with resolution dependent universal thresholds. For each filter, 100 random frequencies are drawn from a uniform distribution on [/8, 7/8]. Tests statistics are computed for each frequency, hence the cut-off value is set to n cut = 17, as suggested by Gottwald and Melbourne (2009). The final statistic for each filter is the median of these 100 single frequency test statistics.

Resende and Zeidan (2008) results
On page 34 Resende and Zeidan's (2008) Table 1 displays the coefficient and the correspondig p-value. We assume that "coefficient" meansα 1 for each currency. However, p-value computation deserves some concern. 3 As FSA argue, the null hypothesis (chaos) includes a wide range of deterministic chaotic processes. Thus, they compute some critical values for just two sample sizes: 380 and 2000 observations, by combining the empirical distribution of the statistics from 250 replications of three of those processes: Logistic, Henon and Lorenz. In Resende and Zeidan (2008), the sample size is 167 observations, hence they should compute the new empirical distribution. Nevertheless, there is no mention to this issue neither to the mentioned procedure to compute the p-values. 4

Our results
The analyzed data, as in Resende and Zeidan (2008), correspond to New York Money Market Survey (NYMMS), regarding expected future exchange rates from 30 traders, at one week horizon, for the British pound (BP), the deutschemark (DM), the yen (JY) and the Swiss franc (SF). The data set comprises 168 observations for each currency. Provided the evidence about unit roots, we compute first differences to render the data stationary. Next, Resende and Zeidan (2008) compute the residuals from fitted ARMA models for each first-differenced currency time series, to calculate FSA test on such residuals. 5 We compute the statistics directly on the residuals series for embedding dimension ranging from 2 to 6. 6  In Resende and Zeidan (2008), there is an inconsistency between their Table 1 results and their conclusions. The computed p-values display evidence of chaos for BP at embedding dimension 5, and statistical evidence of chaos for JY at embedding dimension 6. However, they conclude that there is evidence of chaos for DM and JY. Our results do not confirm their results displayed in the text, either in their Table 1. We just find weak evidence of chaos for JY, for just one embedding dimension as well, but with a rather different p-value: 0.339 vs. 0.886. 7 5 The original residuals have been kindly provided by Rodrigo Zeidan. 6 When fitting an ARMA(p,q) model, the sample size of the residuals is reduced depending on the number of AR lags of the fitted model. This reduction is not accounted for in Resende and Zeidan (2008). 7 We contacted the original authors to reconcile the results, but were unable to deter-Regarding Shintani and Linton (2004) test, the lag (d) and the number of hidden units (r) are jointly selected based on BIC. QS kernel with optimal bandwidth is used for the heteroskedasticity and autocorrelation consistent covariance estimation. Based on the Full estimation version of this test, Table 2, there is no evidence of chaos, just marginally for BP, confirming FSA test results. 8 In addition, we compute the 0-1 test, Table 3. THe conclusions are sharply differente. Hence, with just one exception (JY with C(12)), the test statistics are close to one as they range from 0.9891 to 0.9980. They strongly confirm the hypothesis that fluctuations of exchange rates expectations are, at least, partially chaotic.

Conclusions
In this paper, we have revised and extended Resende and Zeidan (2008) results. We have stressed some important details of the testing process using FSA test. In addition we have applied Shintani and Linton test and the 0-1 test.
Results of FSA and Shintani and Linton test confirm Resende and Zeidan (2008) conclusions, hence exchange rates expectations are not chaotic. However, results with 0-1 test point towards chaotic motion (at least partially, after filtering out noise) in all exchange rates expectations. mine the source of the discrepancies. It should be due to differences in any of the relevant parameters to compute the statistics, since the residuals series used in their work are the same that we have used in this paper.
8 For the estimation based on blocks (Block) and ES, conclusions are identical and are available upon request.