Referential Signaling (White Paper)
Tradescapes and Referential Signaling
A tradescape is a graphical visualization that combines the value of reward-to-pain performance metrics with the science of EM (expectation modeling). You should first familiarize yourself with these concepts in order to better understand this discussion of referential tradescapes.
Managing Abrupt Price Movements
The price movements in an individual security can be abrupt and chaotic, often playing havoc with signaling algorithms that seek to process the order within the time series.
Many signaling algorithms work on the principle of estimating a smooth expectation of price. For a single component signaler, entry and exit signals follow from that smooth estimate changing sign in its slope. A two component signaler might operate on the the difference between two smoothed signals that incorporate different amounts of information, as in the crossover of a fast and slow moving average.
To be more responsive to abrupt events, the obvious solution would be to signal when the current price, with no lag, crosses above or below a smooth estimate of price. This usually works poorly because sharp abrupt price movements can oscillate wildly in both directions, and the process can repeatedly cycle in what traders call whipsaws. Such trades are usually for losses.
One solution to managing this abruptness is to set a band about this smooth estimate creating a price channel. Many signaling systems use a volatility channel, estimation uncertainty, or trading range band around the smooth estimate. The entries and exits are then made using some form of spacing derived from the bands defining the channel. That spacing itself adds lag and inefficiency to the signaling, but it acts to prevent most whipsaws.
A breakout signaler is really nothing more than a price channel that uses an n-day trading range. Since the highest high in n days and lowest low in n days are actual prices, the bands do not require a central smoothed average. Unlike the smoothed methods that have at least a more uniform lag, breakouts distribute across the entire dimension of the sliding window. For a 20-bar window or time horizon on the signaler, the upper price limit could have been set 1 bar ago, the lower limit 20 bars ago, or the reverse could have been true. The upper and lower bands that will have that much variability in terms of lag at each point in time.
In all of the methods discussed, the solution to the chaotic price movements was to make the signaler more responsive. There is an alternative.
Often chaotic price movements are local in the sense that they are specific to that particular security as opposed to the market as a whole. Earnings reports, news specific to this individual company or to it and just a few others, would be examples of external events that perturb an individual security's price adversely in terms of a direct signaling approach.
An alternate approach, and often a quite a viable one, is to do exactly the opposite. Instead of seeking something more responsive, and assuming the abrupt events to be valid and worthy of a swift reaction, we can seek something less responsive. In effect, instead of seeking something that sharpens this vision of movements, we look for something fuzzier, something that is less likely to react.
This tends to work best with stronger entities. To take this approach, the trader must necessarily assume the market is fickle and inefficient, and that it will punish and reward needlessly and well out of line with the security's long term smoothed expectation.
Referential signaling implies that there is a surrogate, an entity upon which the signaling will be based, that will furnish better entries and exits than directly signaling the entity being traded. This can work for the simple reason that the signal target, say it is a broad market index, will move more smoothly than the individual security. If it does, the signaling algorithm will typically have less lag. And the biggest 'if' follows: 'if' there is a close enough correlation between the surrogate and the security, the additional fuzziness will be more than offset by the improved accuracy that comes from the more stable signaling.
It does work. In our experience building trading systems for securities with a strong trending history, we often found that an overall market sentiment filter traded an individual entity more effectively than anything we could signal directly.
If one wishes to take this alternate perspective, one must identify a surrogate, an entity upon which to signal that is closely enough correlated with the behavior of the security. Once that is accomplished, the reference can be used as the data source for generating the entry and exit signals which are then applied to the security of interest. This is referential or surrogate signaling.
Referential signaling is quite distinct from pairs trading where one seeks to exploit an arbitrage differential between two very similar entities. The surrogate may be seemingly unrelated, as in a very specialized company that is traded using the signals from the DJ30.
Referential signaling is a science that is becoming more rather than less prominent as the entities within markets become ever more tightly bound, and even the various global markets are starting to move more in synch. Given that the global economy ever presses forward, it is hard to imagine this aspect of the science diminishing in its usefulness.
Finding a Suitable Signaling Target
This is no simple task. In fact, it is exceedingly challenging. In terms of the mathematics of it, one could say that it is the leap from solving a one dimensional problem to solving a two dimensional problem. If the direct signaling problem consists of a hundred different securities in a candidate pool, the referential signaling problem will consist of a hundred different securities and perhaps a hundred different surrogates. The degree of difficulty increases dramatically. The permutations are now 10,000 instead of 100.
In many cases, a fundamental analyst can point the way. If one is looking at an energy company, one of the energy ETFs or perhaps Exxon-Mobil or Chevron might be viable targets for the referential signaling. If one is looking at a new web-based retail business that might track the overall consumer sentiment, a major index such as the S&P 500 or the Nasdaq might be interesting targets. If it is a tech business that is truly global, perhaps with a focus in Asia, one might try the Asian market indices or another global index from a Western economy that has a high level of exports. In the US, there are the many ETFs that reflect international markets that would allow one to referentially signal a US entity using an international market behavior, and have the same trading days as a US security.
Common sense should prevail. Whatever one selects for a reference needs to make sense. Totally unrelated entities can be correlated purely by random chance. In our work, we have largely restricted the references to major market indices, world indices, international ETFs, sector indices, and sector ETFs.
When using this approach for advanced signaling, bear in mind that the added complexity opens up another dimension for random chance as well. It is possible that historically, even across a decade, one might find that a candy company's stock is signaled to perfection by using a defense contractor's stock price as a reference. If you look at enough of a matrix, these random monkeys will appear. Again, to a fundamental analyst it will be second nature. Use a semiconductor index as a surrogate for a company that makes semiconductors, a transportation index for a company that produces truck engines, and so forth.
To make this very difficult process tenable, we've extended tradescapes to handle referential signaling. A tradescape is usually constructed based on signaling the traded entity directly, but there is no reason why it cannot be constructed referentially. The EM signals can be drawn from any surrogate and then applied to the entity one wishes to trade. We call the end product of this type of visualization a Referential Tradescape.
There are two ways to do a referential tradescape analysis on a set of securities. One can specify a trading target from one of the entities in the data or chart. All of the other entities are treated as references. All of the tradescapes will reflect the trading of one entity, with many references. This is the more common approach, and it can make the search for a reference far simpler than it would otherwise be.
The other approach is to specify a reference from one of the securities in the set. The tradescapes will then reflect what this particular reference can accomplish as a surrogate across the other securities in the chart. This is useful if one is building a portfolio that is being signaled on an overall market index. One would want every entity in that portfolio to be favorably signaled by that single reference.
A First Example
It is well known that trading SPY, or ES, or any SP500 related entity is very difficult. The SPY tradescape bears this out.
There is some function at low lags and at both very fast and very slow signaling, but not much between. Since we know many world markets follow the US, let's look at SPY as a surrogate for many of the international market ETFs sold in the US. The referential tradescapes are shown on a grid of contour plots.
The SPY direct tradescape is in the first plot. The remaining plots are for DIA, EWZ, EWA, EWM, EWS, EWT, EWG, EWH, EWC, EWW, EWY, EWJ, and GLD using SPY as the signal reference. If you look closely, you will see little difference between SPY and trading DIA using SPY as the signaling target. That would be expected. In terms of the various countries, we realize less return (using RRt reward-to-pain) when trading the Taiwanese (EWT), Hong-Kong (EWH), and Japanese (EWJ) ETFs using SPY as the signaler than we see trading SPY directly. On the other hand, Canada (EWC) and Mexico (EWW) are stronger as is Singapore (EWS) and Malaysia (EWM). GLD is also traded effectively using SPY, at least for the 7.5 yrs of data available (all of the other use 10 yrs).
With a thirty second analysis, no work yet in terms of constructing any trading system, we know that we can use SPY as a potential surrogate in certain places and not others, and we know the lag tolerance any trading system will need to have and about what time horizon or trade density to use in order to be successful. There are over 10,000 EM-type backtests in the referential tradescape grid above.
Let's now try the referential process in the opposite direction. Let's see if any of the entities here work effectively for trading EWA, the Australian ETF. We know the Australian economy is strongly linked to the Asian markets.
The fourth tradescape is the directly signaled EWA. None of the references used here show much value, although EWT, the Taiwanese ETF does offer a good measure of fast signaling function if one could construct an accurate system with low lag.
While the process does require some fundamental insight, it is possible to know within a relatively short time what is possible in terms of referential signaling. Bear in mind that the first matrix looked at the value of trading the international ETFs using SPY as a surrogate and the second matrix looked at trading EWA with a variety of surrogates. In the first case, we are looking for an entity that likes to signal on SPY. In the second we are looking for surrogates that trade one specific entity, EWA, effectively. We would expect the tradescapes in the second matrix to be more self-similar. Still, there are profound differences in both cases.
A referential tradescape is also a good way to test certain fundamental ideas. For example, Australia is the world's second largest gold producer. One might expect the overall movements in the price of gold and that of the Australian markets to track somewhat closely. In actual fact, except in the very fuzziest long term sense, there is no referential trading relationship at all. At any reasonable signaling density, there is more pain than return, and this is with the 100% accuracy in the tradescape's signals. Real-world signals will be even worse.
A Second Example – Good Can Get Better
In the first example, we looked at entities that are generally considered quite difficult to trade. In this example, we'll return to AAPL, an entity shown to be very forgiving historically, and we'll see if we can find any benefit from a referential signaling.
Immediately, one observes an immense difference in the referential tradescapes. The directly signaled AAPL tradescape is the first plot. There is great trading function through a lag fraction of 1 if the right trade density is chosen. We look at whether this very happy state of affairs can be further improved upon using the following potential surrogates, the DJ30, GLD, the SP500, the Hang Seng, the IIX Internet index, the Nasdaq 100, the Stockholm index, SPY, and the XAX, the Amex Composite.
Despite the fact that both gold and Apple Computer have shown remarkable growth in the past decade, there is little in common referentially speaking.
The SP500 and SPY tradescapes, the DJ tradescape, and the Nasdaq tradescape show some benefit at very high lags. This would likely benefit only weaker higher-lag signaling systems.
There is a benefit from using the IIX Internet index or the rather export oriented Stockholm index. Properly designed, a well-crafted system may see a benefit from signaling on these entities. Note how the higher RRt, colored in blue, extends to higher lag fractions with these two surrogates.
If we look at the 3D surface plot format for the tradescapes, we see that the Internet and the Stockholm index behave very differently from one other. Both produce a "fuzzier" response surface, and both produce a larger, flatter blue plateau that extends to greater lags. The OMXSPI offers considerably more reward to pain at every lag and time horizon in the tradescape. That is impressive. Note that this tradescape never decays to the 1.0 RRt where reward and pain are equal, not even at the highest information lengths and the highest lags.
If the issue is robustness, or accommodating a signaling system that can't quite reach those lovely blue and magenta peaks in the direct AAPL tradescape, these referential signals may offer a viable alternative to direct signaling.
Again we must apply common sense in any referential design. Here both make sense. The IIX is closely linked to AAPL's businesses, and the Stockholm market is one of the most export-oriented of the first world markets, very fitting AAPL's global business.
Evaluating a Real-World Referential Trading Signal
The same approach for evaluating signals on a direct tradescape is used for a referential tradescape. You must construct a binary signal of 1's and 0's based on the reference instead of the entity to be traded.