Skip to content

Empirical Quantiles and Proxy Selection

October 3, 2011

The previous post on Proxy / Cross Hedging left open an important question: how to choose an appropriate hedge instrument, especially amongst several alternatives. Prior to diving into full exploratory data analysis for proxy hedging (in a forthcoming post), this question is worth careful consideration.

One way to approach this problem is via visualization of empirical quantiles.

Consider the defining characteristic of an ideal proxy for neutralizing price: the difference in weighted prices of u and h is zero for arbitrarily few samples (where sampling frequency is assumed to be no larger than the rebalancing period). The hard part is the requirement for arbitrarily few samples; more formally, the \epsilon time series for ideal must satisfy:

   \sum\limits_{i=1}^n \epsilon_i = 0 : \forall n \ge 1

Rather than the much looser requirement that \epsilon \rightarrow 0 in \mathbb{P}.

Contrast this requirement with how statistics normally approaches this problem, using techniques essentially derived from moment generating functions. Measure theory would incorporate almost sure equality. Yet, the ideal proxy conceptually demands more than both; it demands equality in empirical distribution. Hence, any prospective solution must be framed in terms of both empirical samples and the entire distribution of both u and h.

Thus, the question boils down to evaluating joint comparison and equality of empirical distributions. Doing this using raw prices is unnatural, as the two instruments use different metrics. Instead, transform into linear returns for both (not log) and convert into a single returns metric. Thus, reformulating the proxy hedge model, using the discrete derivative operator D:

   D(u) - \beta D(h) = \delta

From which the empirical distributions for D(u) and D(h) can be generated, and ideal proxy reformulated as:

   \prod\limits_{i=1}^n (1+\delta_i) - 1 = 0 : \forall n \ge 1

When this equality holds, the empirical distributions are strictly equal, subject to the linear \beta scaling. Now, the question boils down to evaluating distributional equality while ignoring \beta. For this, we appeal to quantile analysis (previously discussed in the context of stability, and nicely exemplified in research by Koenker). Specifically, use of quantile-quantile (QQ) plot.

Consider the QQ for both empirical distributions D(u) and D(h), rather than traditional comparison between theoretical and empirical. Thus, the plot illustrates the desired visualization of empirical distributional similarity. Most importantly, the \beta scaling disappears due to measurement in units of quantile. Further, this approach is non-parametric, obviating the need to posit distributional assumptions on either instrument.

From this QQ, the evaluation of proxy optimality is made visible:

  • Optimality: ideal proxy hedge will have a qq line which passes through all observations
  • Divergence: divergence from optimality is manifest by observations away from the quantile line, whose respective distance away from the (0,0) origin reflects its corresponding frequency of occurrence

For example, consider the following proxy QQ plot:
Empirical Proxy QQ Plot

Which plots the empirical quantiles of daily returns for a well-known high tech company (CRM) versus the Nasdaq-100 Index (QQQ). The solid dashed line is the (0,1) line, reflecting distributional equality. The dotted blue and red ellipses are location dispersion ellipsoids for the two distributions (following Meucci 2010).

Thus, visual inspection of this plot quickly rules QQQ out as an effective proxy:

  • Observations across all quantiles are off the line, indicating ineffectiveness at all quantiles
  • Dispersion ellipsoids are significantly rotated, indicating the principal component is strong misaligned
  • Outliers in the tails demonstrate both extreme divergence and non-linearity

Thus, not only is QQQ a bad proxy, it is arguably a dangerous proxy from tail risk standpoint: big moves in CRM will be badly underhedged by QQQ. So much for the standard hedging claims of QQQ futures.

What is particularly notable is these observations are drawn from a single plot. As the next post on full exploratory data analysis for proxy hedging will illustrate, this is further notable as numerous other more popular visualizations illustrate the contrary conclusion.

For readers interested in further study, Doyle (2010) discusses several further mathematical nuances of empirical QQ plots.

R code to generate the proxy QQ plot:

proxyQQPlot <- function(p)
  # Plot proxy QQ plot, including empirical quantiles and location dispersion ellipsoids.
  # Args:
  #   p: xts of instrument price data, including valid colnames
  # Returns: None

  p1ROC <- ROC(p[,1], type="discrete", na.pad=FALSE)
  p2ROC <- ROC(p[,2], type="discrete", na.pad=FALSE)

  qqplot(coredata(p2ROC), coredata(p1ROC), xlab=paste(colnames(p)[2], "Returns Quantiles"), ylab=paste(colnames(p)[1], "Returns Quantiles"), main="Empirical Returns QQ-Plot")
  d <- dataEllipse(as.vector(coredata(p2ROC)),as.vector(coredata(p1ROC)),draw=FALSE)
  lines(d[[1]], col=colors[2], lty=3)
  lines(d[[2]], col=colors[3], lty=3)
8 Comments leave one →
  1. Breno Neri permalink
    October 4, 2011 6:12 am

    Sorry, but your first equation, \sum\limits_{i=1}^n \epsilon_i = 0, \forall n \ge 1 , implies that \epsilon_i = 0, \forall i \ge 1 🙂

    Trivial to prove that by induction.

    • quantivity permalink*
      October 4, 2011 8:23 am

      @Breno: thanks for your comment. Your induction proof is indeed correct, and is the intended dynamics for the ideal \epsilon. I will edit the text to clarify this intent.

  2. John Hall permalink
    October 4, 2011 7:04 am

    Why not just minimize the 50% CVaR deviation while fixing the weight of the stock in focus at 1 and imposing a cardinality constraint of 2. You could do that without needing to resort to a formal optimizer.

    • quantivity permalink*
      October 4, 2011 12:34 pm

      @John: not sure I understand your suggestion; can you please clarify which deviation (I guess \epsilon) you are proposing to minimize and which variable is the cardinality constraint being imposed (I guess hedge portfolio size)?


  1. Empirical Copulas and Hedge Basis Risk « Quantivity
  2. Finanzas 101: Proxy Hedging « Quantitative Finance Club
  3. Proxy / Cross Hedging « Quantivity
  4. Exploratory Hedge Analysis « Quantivity

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s

%d bloggers like this: