Extracted main text — title through conclusion, appendix excluded. This is what our citation measures are computed over, published so the extraction can be checked by eye.
29,684 characters · 12 sections · 19 citation commands
NAPLES;Mining the lead-lag Relationship from Non-synchronous and High-frequency Data
The lead-lag effect is a phenomenon wherein the change of one time series affects another time series with a delay. This effect appears in various fields from financial markets to control engineering. In particular, the lead-lag effect in financial markets has been thoroughly investigated because it is a crucial factor in investment, and the knowledge of the size of the lag can potentially help investors make a profit without taking any risk. Discovering lead-lag effects is vital because high-frequency trading uses high-speed communication, and the high-performance computers used in high-frequency trading can make profits by using unapparent lead-lag effects in millisecond and nanosecond timescales Biais2015,Wang2015,Martinez2012. Moreover, high-frequency trading also makes the market more "efficient." In high-frequency trading, the lead-lag effect does not usually persist for a long time because market investors identify such effects, and are often quick to take advantage of such asymmetry in the market CHABOUD2014,Case2011,Bhaskkar2008. Therefore, to make a profit, financial analysts must be quick and accurate in estimating the lead-lag effect.
To detect the lead-lag effect, we formulate the following maximization problem(eq.(ref)). Let $X_t,Y_t$ be asset prices, and we assume that $X_t$ is observed at $s_1 < s_2 < \cdots < s_n$ and $Y_t$ is observed at $t_1 < t_2 < \cdots < t_m$. We say that $X_t$ leads $Y_t$ with a lead-lag parameter $\theta$ if the correlation between $X_t$ and $Y_{t+\theta}$ is stronger than the correlation between $X_t$ and $Y_{t+\theta'}$ for any $\theta' \neq \theta$. We find the maximum lead-lag parameter $\hat{\theta}$ as follows:
where $\hat{\mathrm{Corr}}(\{X_{s_i}\}_{i=1}^n,\{Y_{t_j}\}_{j=1}^m)$ is a correlation estimator for the observed time series, and $\Theta$ is a candidate of the lead-lag parameters.
In estimating the lead-lag effect in high-frequency trading, there are three major challenges: (1) Non-synchronicity of timestamps (2) Computational complexity (3) Time-variance of the lead-lag effect. (1) By the non-synchronicity of timestamps, we mean that the price of each asset is observed independently and irregularly. In high-frequency trading, price movement occurs irregularly since transactions happen asynchronously and independently of each asset. Non-synchronicity makes it difficult for us to estimate the correlation. Consequently, it does not allow us to estimate a lead-lag effect by simple methods. Hoffman, Rosenbaum, & Yoshida(2013) proposed a solution to this problem using an estimator that has mathematical convergence Hoffmann2013. Moreover, various theoretical analyses Chiba2019,Hayashi2017,Hayashi2018, and empirical analyses Alsayed2014,Bollen2016,Ceron2016,HuthAbergel2011 have practically applied the method. (2) Computational complexity is a significant challenge in high-frequency trading, for which relevant information is becoming increasingly available. The number of assets and lag candidates also increase, along with the complexity of the relationships between them. If the length of the time series is $T$, the number of assets is $N$, and the number of lag candidates is $L$, the current extent of the calculation is in the order of $O(T^2 N^2 L)$. A calculation algorithm better than this order has not been proposed.
(3) The time-variance of the lead-lag parameter is another critical factor that has not been explored in depth. The time-variance of the lead-lag relationships means that some lead-lag effects occur temporally. Most existing lead-lag estimation methods assume that the lead-lag parameter $\theta$ is constant; therefore, they can be applied only to limited situations. This time-varying nature of the lead-lag effect has been observed empirically Huth2014. In this paper, we propose a new estimator of lead-lag effects to solve these three problems. This method can not only be performed using a non-synchronously observed set of time series, but it can also be performed with a computational cost that is linear to the number of observations $O(N)$. The main contributions of this paper are as follows.
We denote the interval of the numbers between $a \in \mathbb{R}$ and $b \in \mathbb{R}$ as follows: $[a,b]=\{x | a\le x \le b\}$ ,$]a,b] = \{x| a < x \le b\}$, $[a,b[\ = \{x | a \le x < b\}$, $]a,b[ \ = \{x| a<x<b\}$. $a\vee b$ and $a \wedge b$ stand for operations $\max\{a,b\}$ and $\min\{a,b\}$, respectively. $1_A$ denotes an indicator function which is $1$ when the condition $A$ is true and $0$ when $A$ is false. $\theta$ denotes the lead-lag delay parameter between two time series, and $\rho$ denotes the correlation between the two time series.
We formulate the lead-lag estimation problem mathematically. First, we define the notion of the lead-lag effects and non-synchronous observations. Let $X_t,Y_t$ be two stochastic processes (e.g., asset returns). Assume that $X_t$ is observed at $0=s_1<s_2<\cdots<s_n=T$ and $Y_t$ is observed at $0=t_1<t_2<\cdots<t_m=T$.
Our main problem is to infer the lead-lag effect from non-synchronously observed data.
Many studies, such as KAWALLER1987 and Lo1990, showed the existence of the lead-lag effect for synchronously observed time series. Because high-frequency trading has become the standard trading method, the importance of modeling and analysis of non-synchronously observed data is increasing both in theory and in practice. Hayashi and Yoshida Hayashi2005 proposed the following covariance estimator.
They proved the convergence of this estimator when $X_t,Y_t$ are in Brownian motion Hayashi2005.
On the other hand, Hoffman et al. Hoffmann2013 showed that the covariance estimator with a transferred time series $U^n(\tilde{\theta})$ can estimate the lead-lag effect when $X_t,Y_t$ are in Brownian motion.
\[ U^n(\tilde{\theta}) : = 1 _ { \tilde { \theta } \geq 0 } \mathrm{HY}(X_{s-\tilde{\theta}},Y_t,\{s_i-\tilde{\theta}\}_{i=1}^n,\{t_j\}_{j=1}^m) + \] \[ 1 _ { \tilde { \theta } < 0 } \mathrm{HY}(X_{s+\tilde{\theta}},Y_t,\{s_i+\tilde{\theta}\}_{i=1}^n,\{t_j\}_{j=1}^m). \]
Chiba Chiba2019 extended this estimator to the case when $X_t,Y_t$ are partially in Brownian motion. Hayashi and Koike Hayashi2017 improved the estimator with a Fourier transform and Hayashi and Koike Hayashi2018 improved it with a Wavelet transform.
In contrast, Dobreva and Schaumburgb DobrevaSchaumburgb2017 introduced the following lead-lag estimation methods.
We can estimate the lead-lag effect with the Dobreva-Schaumburgb index in the same way as Hoffmann2013. Its statistical properties are not apparent, but empirical studies show its effectiveness.
We define the index $R(t)$ for the lead-lag effect estimation, which is the main method of this paper.
Before defining our lead-lag estimator, we prepare some notations. Assume that we observe $\{Z_{u_k}\}_{k=1}^n \in \{\{X_{s_i}\}_i,\{Y_{t_j}\}_j\}$. Let $r^{(Z)}_{u_k}$ as the logarithmic returns of assets $Z$, $ r^{(Z)}_{u_k}=\log (Z_{u_{k}}/Z_{u_{k-1}}),\ $. Let $b^{(Z)}_{u_k}$ be the sign of $r^{(Z)}_{u_k}$, $ b^{(Z)}_{u_k}=\mathrm{sign} \{r^{(Z)}_{u_k}\}$. Then we denote $\hat{Z_t}$ as the cumulative sum of the signs of returns $b^{(Z)}_{u_k}$,$ \hat{Z_t} = \sum_{k=1}^n 1_{u_k < t} b_{u_k}^{(Z)},\ \ $
This definition looks complicated but is based on a simple idea. If $X$ is leading $Y$, trading $Y$ following the price movement of $X$ makes a profit or a loss. $R(t)$ is an index that expresses this idea numerically. The first term \[ b^{(X)}_{s_ {i}} \hat{Y}_{s_{i + 1}} -b ^ {(X)}_{s_ {i}} \hat{Y}_{ s_ {i}} \] is the return of the trading strategy which buys $\hat{Y}$ at $s_i$ and sells $\hat{Y}$ at $s_{i+1}$ when $X$’s price rises at time $s_i$ and does the opposite when the price of $X$ falls.
The second term \[ b^{(Y)}_{t_{j}} \hat{X}_{t_{j + 1}}-b^{(Y)}_{t_{j}} \hat{X}_{t_{j}} \] is the return of the trading strategy which buys $\hat{X}$ at $t_j$ and sells $\hat{X}$ at $t_{j+1}$ when $Y$’s price rises at time $t_j$ and does the opposite when the price of $Y$ falls. Figure (ref) is the illustration of our idea.
This method solves the three problems described above. (1) For the issue of non-synchronicity, we can use this method for time series of non-synchronous observations. (2) To solve computational complexity, our estimator is simply a sum of $N$ terms; therefore, we can calculate this estimator with the order $O(N)$. (3) To address the time variance of the lead-lag effect, our estimator is a return of a trading strategy so that we can calculate the lead-lag parameter at each time.
To clarify the mathematical properties of $R(t)$, we present the following theorem.
We show the general case of this theorem and its proof in the Appendix. Thus, we can estimate the lead-lag parameter $\theta$ by calculating
We interpret the main Theorem (ref). In Equation (ref), we consider a situation where we want to check if the lead-lag effect exists for a given $X_t$ and $Y_{t+\alpha}$. Figure (ref) shows that our estimator achieves the maximum value at $\Delta$ and shows that we can estimate the lead-lag parameters by using argmax like in Equation (ref). We show the meaning of each point in Figure (ref). At point C, there is no lead-lag effect between $X$ and $Y$, so there is no profit if one trades $Y$ according to the price movement of $X$. At point D, there is a lag equal to the trading interval, so we can always make a profit by trading $Y$ $\Delta$ seconds after $X$ moves. At point B, we perform the opposite trade; thus, a loss occurs. On the other hand, at point E and A, if we trade two periods after the price movement of X, we cannot make a profit because $X$ and $Y$ are in Brownian motion, and their increments are independent. In other words, $X_{t+\Delta} - X_t$ and $X_{t+2\Delta} - X_{t+\Delta}$ are independent.
In this section, we experiment with simulated data and real market data to show the effectiveness of our proposed method. We compare our method with the Hoffman-Rosenbaum-Yoshida method and the Dobreva-Schaumburgb methodHayashi2018.
We use the following paired geometric Brownian motion in the experiment. \[ \left\{
\right. \]
We apply the following parameters: the initial value of each asset $ x_0 = y_0 = 100$, volatility of each asset $ \sigma_1 = \sigma_2 = 1.0, \ $, lead-lag parameter $ \theta = 10 $, correlation $ \rho = 0.9 $, candidates of lead-lag parameter $\Theta= \{-100, -99, \cdots, 99,100 \}$. We observed $X_t$ and $Y_t$ at $\{s_i\}$ and $\{t_j\}$ and $s_i - s_{i-1} \sim N(10,2)$, $t_j - t_{j-1} \sim N(10,2)$. We investigated how long each algorithm takes to converge by changing $T$, which is the maximum value of observation time and we use $T \in \{10^{2.5}, 10^{2.6}, \cdots, 10^{5.0} \} $. We evaluated the Mean Absolute Error of the estimation and performed the experiments 1000 times generating random numbers, $n=1000$ times. Let $ \hat{\theta}_i $ be the lead-lag we estimated by each algorithm in the $i$-th experiment and let the real lead-lag be $\theta_i $, then we calculated the average of the absolute errors $ \frac{1}{n} \sum_ {i = 1}^n | \hat{\theta}_i-\theta_i |$. Figure (ref) shows the log of the Mean Absolute Error for each $\log_{10} T$ and that our method converges best to the real value $\theta$.
In this section, we demonstrate that our method can estimate lead-lag effects with actual market data. We estimated lead-lag parameters with actual foreign exchange market data. We chose two currency pairs AUD / USD (Australian Dollar / United States Dollar) and NZD / USD (New Zealand Dollar / United States Dollar), which are geographically close and are categorized as similar assets for international investors. We used the Tick-by-Tick data from January 1, 2019, to September 30, 2019, which were downloaded from Dukascopy Historical Data Feed \footnote{\url{https://www.dukascopy.com/}}. We used "Economic calendar Investing.com Forex (2011-2019)" for the list of important announcements in Australia and New Zealand \footnote{https://www.kaggle.com/devorvant/economic-calendar}. Figure (ref) and Figure (ref) illustrate the lead-lag parameter before and after the important announcements. Moreover, the insights from these figures are as follows:
\if0
\fi
We investigated the estimation problems of the lead-lag effect in high-frequency data. Mainly, we positioned the non-synchronous observation, computational complexity, and time variation of the lead-lag as critical issues. We proposed NAPLES to solve these issues. Our method’s estimation was with low computational complexity. The experiment using simulated data showed that our method captures the lead-lag more accurately than the existing methods do. The experiment using high-frequency foreign exchange market data showed that there was a lead-lag effect at the time of important announcements.