US Dollar Index - DX

[DeepState]

FYI.

Just rebuilding some stuff. Here's some basic results on one widely known strategy. Checking (confirming, really) for evidence of momentum in relative performance. Currencies over-react in the near term for various reasons. Over longer timeframes the behavior is different. Currency is not like other asset classes and does not impound information which is discounted far more quickly in other asset types.

A very basic first step to checking if a concept works is to generate 'signal' and check to see if it has predictive power. That all sounds very obvious. However, the way this is done from the world I came from is not along the time dimension for a single currency. That is, not thinking in terms of a single currency pair and looking for turning points through time. Instead it compares their merits across a set of choices at a given time. The methods are different and so is the trading and portfolio management.

So here are some results on a very simple signal based on the returns of a set of currencies vs USD for one fixed period (eg. like 3 months). It doesn't get much simpler. The currencies are the majors per prior entry. To check for implementation we check for the decay profile (what happens if you are slow to implement). We split the time periods etc.. First, we check for rationale. The tests actually come much later.

There is strong evidence of a viable signal on just this simple idea alone. There are many steps to go from here before this baby is ready for the track. But, it has form.

Data is 20 years to today for six currency pairs alone (with more, you can expect even stronger outcomes). Daily data. Signal is just historical 3 month return vs USD. How simple! There are virtually no moving parts to fiddle and hence data mining risk is way down. You can write the formula on a pea by hand. A good test.

What is being analysed are the ranks of the signal and the ranks of the outcome. This controls for a lot of issues related to risk management or the time being. The decays show how the performance of the strategy varies as you delay implementation from overnight to 20 business days. Notice how it rises before falling...that's because currencies over-react in the near term. It is actually better to let things slide a bit. Perhaps there is room here for T/A type analysis to finesse it.

Results are strongly statistically significant. Analysis of the last 10 years is borderline statistically significant and the shape of the decay curve is basically the same.

WARNING: If you choose to build a portfolio on this idea without some very decent risk management ability any apparent insight will be strongly eroded. Also, this is not the same as saying just because a currency pair has moved in a direction that it will keep doing so.

Because there are no fundamentals used as a direct input, you could possibly label it as a form of T/A for this predictor. Generally, this type of process would not be called technical analysis in the industry. For me, if it (more likely something different) survives further refinements, it would form one part of a wider set of strategies so the whole suite would not be said to be pure T/A. A mix of ideas helps with robustness. Momentum can be slow to turn or, no surprise, overrun.

Here is an idea which has good rationale before examination of outcomes, is formed by a method that is almost impossible to simplify, is strongly statistically significant, easily covers frictions (what frictions?), and can be implemented due to adequate time allowance for trading prior to decay.





If there are others out there working on anything in this area....

 

[DeepState]

FYI

The results below follow on from the prior posts and are for interest rate differentials. Carry. The difference in interest rates between two countries is a predictor of the movement of the exchange rate between them. The higher interest rate country is expected to appreciate. This is termed a violation of the uncovered interest rate parity relationship. It is probably the single biggest violation of an arbitrage related financial relationship in terms of money transfer between what is predicted in pure theory sense (interest rate parity should provide an unbiased estimate of currency movement) and what actually happens (the opposite).

The charts below are for the 10 year period to Friday.

Key take-outs:

This is a super-powerful (the columns are high) slow burn idea (the columns are flat) but hit a sudden stop at the GFC. In terms of the actual relationship between the predictors and the outcomes on a rank-to-rank basis, there was a hit, but it has been relatively flat line since then.

A key reason why the idea hit a stop was that interest rate differentials converged post GFC. If you excluded Australia from the stats, the chart showing the standard deviation of the cross section would be much lower. There was virtually no interest rate differential to profit from hence the predictor is very weak....that might all be changing with the normalization of policy in the UK and US. After a period of dormancy, carry is likely to be returning and it is visible in the stats.

The underlying drivers of carry and the risks to it can be ascertained via examination of national accounts and other financial market information. You do not (have to) simply rely on the raw figure. That said, the raw figure works when there is opportunity to do so. The opportunity is returning.








If anyone out there is a discretionary FX trader/investor who considers such matters and underlying economic data, it would be good to hear from you. My trade horizon is measured more at the month(s) level than intraday. My assessment window is measured in years. These are slow moving ideas.

This stuff works. It provides a basis from which discretionary analysis can be used to confirm or deny the validity of the signal. This is essentially the way to ascertain whether the currencies are pushing their luck.

The next step is to produce risk management estimates to assist with position sizing across a portfolio of FX exposures. I intend to combine GARCH family estimates on individual PCA components, recombined into a covariance matrix and cross check the outcomes with 21d VaR analysis. The final portfolio will be strongly informed by these but is ultimately discretionary.

 

[DeepState]

....and as if to confirm the assertions for the prior post, here are the figures for 10 year bond interest rate differentials. Although the target of QE of various forms, these are less directly controlled by the CBs and hence their dispersion was not compressed by as much. The results were better than the ones I showed before (which were for 3 month maturity).

 

[rimtas]

RY, just a simple question- what is an outcome of your observations and analysis, where the US Dollar is going medium/Intermediate term?
Because any analysis should eventually generate simple buy/sell signals depending on the time frame analysed.
As the worlds reserve currency, US Dollar is the leading indicator to everything-if it rises, it indicate a deflationary pressure on all markets. By determining a probability of the future trends in USD we can anticipate how bond and stock markets will behave.

Thanks,

 

[DeepState]

RY, just a simple question- what is an outcome of your observations and analysis, where the US Dollar is going medium/Intermediate term?
Because any analysis should eventually generate simple buy/sell signals depending on the time frame analysed.
As the worlds reserve currency, US Dollar is the leading indicator to everything-if it rises, it indicate a deflationary pressure on all markets. By determining a probability of the future trends in USD we can anticipate how bond and stock markets will behave.

Thanks,

Here's a representation of a simple 3 month momentum signal and 10 year bond interest rate differential signal as at current day.

Essentially, the model would suggest buying GBP and CAD vs US and funding it from JPY vs USD. The rest are not high conviction positions.

One of the features of this approach is that it regards the USD as the price of each currency and then tries to figure out how the price of each currency will change against each other. Everything is relative in that way and it will not generate an absolute Buy or Sell USD vs Rest of World. It generates a portfolio of pair trades amongst the major currency pairs vs USD.

This is a pair-trade model. Deep inside it is an attempt to figure out what the best exposures are to 15 possible pairwise currencies (ie. AUDGBP, GBPCAD, JPYEUR...) and to achieve these via USD denominated positions. In other words, what it is really saying is it likes GBP and CAD vs JPY. Along the way, you can make an inference for what they think about the USD - but that's not always the case.

Your simple and clean question has been met with a complex response. Still, I hope it adds a little to your adjudications.

 

[rimtas]

Thanks RY, the answer satisfies me

 

[DeepState]

The next step is to produce risk management estimates to assist with position sizing across a portfolio of FX exposures. I intend to combine GARCH family estimates on individual PCA components, recombined into a covariance matrix and cross check the outcomes with 21d VaR analysis. The final portfolio will be strongly informed by these but is ultimately discretionary.

Here's what it looks like:

--





--
Here is what it means:

For a data window length of 250 business days going back from yesterday, risk for a currency basket is calculated.
The currency positions are a USD 1m LONG AUDUSD and balancing SHORT CADUSD. In other words the basket is LONG AUDCAD for USD 1m.

Two risk measures have been used and two time horizons have been examined for each. One is a Horizon period. I have chosen this to be 21 business days, or basically a calendar month. The other is daily.

The first risk method is GARCH (1,1) PCA. In essence, it breaks down the performance of the currencies under consideration into underlying statistical drivers. Each of these is then individually examined for risk characteristics and forecast using a method called Generalised Autoregressive Conditional Heteroskedasticity. This class of time series models is seen to be a solid estimator of future volatility given observed volatility. These forecast risks are then recombined to form a representation of the risk/correlation relationship between the currencies for the next 21 days (this number can be changed). The outcome is the standard deviation of dollars at risk based on single standard deviation over the horizon period (~25k). You can also infer what that means for a daily level of risk (~5k). Over the relevant period, dollars made or lost are expected to be within these figures 67% of the time. You can double the figures and then you would expect that, 95% of the time, your P&L will be within that band.

To cross check this, Value-at-Risk has also been calculated based on data over the selected window. This looks at the returns which this portfolio would have achieved in history. Daily and rolling 21 days results are calculated. The figure of interest is what the worst 5% of these observations was. This is known as VaR(5) for each series. You can change the threshold to whatever you like, but 5% and 1% are the standards. Over the window, 5% of rolling 21 day outcomes were worse than ~38k. Over a 1 day period, the corresponding figure is ~7k.

The GARCH PCA is showing a higher risk level once the probabilities are brought to a common basis. That is because it takes into account recent risk spikes in a more sensitive fashion. Currencies have been more volatile over the recent period than over the year on average. Risk spikes tend to persist for a while. Hence the GARCH PCA method is estimating higher than usual risk. In contrast the VaR analysis just looks at things over the window and does not make those types of adjustments. The benefit of VaR is that is takes into account the shape of the return distribution in a way that GARCH PCA cannot.

Both measures are informative when deploying and analyzing portfolio risk.

In combination, this tool will assist with determining how to size positions given a particular level of risk appetite. By using it in a different way, it will also help with determining the composition of that risk (how to size different currency pairs when there are many moving parts).

These are the basic tools that are used in the professional market for simple situations like this. You can add a lot more bells and whistles. In my case, I'm not running anywhere near as tight to require much building on top of this in a time series statistical sense. Positions will also be supplemented by stops. Other risk management takes the form of fundamental analysis of the economies in question and skews inherent in the currency options market.

Once again, if there is anyone out there doing something of this nature, dialogue would be welcome.