forex-scalping.com

Correlation tables and the horizon problem

A correlation table says EURUSD and GBPUSD correlate at 0.85. A scalper takes it as a fact. It isn't one.

Correlation depends on the horizon and the period. Measured on daily data over a year, the figure might be 0.85. Measured on one-second data over an hour, it can be 0.3, or negative for brief stretches, because the two pairs update at different moments and react to different news.

It also changes over time. During a eurozone crisis, the euro and sterling can decouple entirely. During a dollar panic, everything correlates.

For scalping, the relevant correlation is the one that holds on the timeframe of the trade, right now. A rolling calculation over the last ten or thirty minutes, on the frequency traded, tells that.

Using stale numbers leads to wrong assumptions about diversification. Two trades believed to be uncorrelated may be effectively the same bet, or two believed to be hedged may drift apart.

Even a rough rolling measure beats a table from a website. And it should be recalculated often, since it doesn't stay still.

At tick frequency, how does the rolling correlation of two pairs compare with the daily figure?

Correlation tables show how closely pairs move together, but the result depends heavily on the time horizon used. Correlations calculated from daily data over a year can look very different from those calculated from five minute data over a week. Two pairs might have a strong positive correlation on daily data but weak or even negative correlation on short intraday periods. For scalpers, the intraday correlation is what matters, since that is the horizon over which trades are held. Using a daily correlation table to manage intraday risk can give a misleading picture of how positions will interact.

I found out how unreliable a daily correlation figure is when I ran EURUSD and GBPUSD together on M1 and thought I had a hedge. The table said 0.85. When I calculated a rolling correlation on one-minute returns over the previous thirty minutes, I saw it swing between 0.2 and 0.7 over a single morning, and on one data release it briefly went negative. My two positions that I considered a hedge were doing completely different things.

Now I compute it in a small script that runs every minute, using the last 45 one-minute returns. If the number is above 0.6, I treat both positions as one trade and halve the size on the second. If it's below 0.3, I count them as separate risks.

One more thing I learned is that the sampling interval matters a great deal. With ticks the figure looked much lower than with one-minute bars, since the pairs update at slightly different moments. One-minute bars were the best compromise for me.