Referential Sentimentscapes (White Paper)
A referential sentimentscape studies the benefit of simultaneously using surrogate signaling and a sentiment filter to generate the trading signals.
There are four basic types of tradescape technologies.
(1) For directly signaling an entity to be traded, we have tradescapes.
(2) For signaling indirectly using a surrogate, we have referential tradescapes.
(3) For adding a sentiment signal to a primary signal that operates directly on the traded entity, we have sentimentscapes.
(4) For sentiment-augmented signaling using a surrogate entity for the primary signal, we have referential sentimentscapes.
Even with the efficiency of information possible via tradescapes, there is an upper limit to what is practical in terms of signal system design. From our perspective, referential sentimentscapes represents that upper bound.
For a referential sentimentscape to be grounded in reality, there should be two independent observations. One should pertain to a signal surrogate that makes sense fundamentally, the other to a market sentiment target that has a sound fit with the entity being traded.
Let's say we want to trade an entity that is difficult to signal directly. For our example, let's assume that is the Australian ETF, EWA, although it could easily be a thousand different widely traded entities that are not easy to signal directly.
Let's start by looking at a long-side 10-yr tradescape of EWA:
This tradescape is a compact picture of the performance that can be attained by an accurate but lagged signal taken directly from the EWA price curve. The news here is generally good. The tradescape shown here sets the lower point of the gradient to an RRt of 1.0 meaning that there is more reward than pain, as defined by robust trend and retracement, everywhere there is a color other than the gray background.
Let us assume that we must use a fairly simple signaling system as opposed to a more elaborate one, and as such we cannot realistically expect to see lower than about a 0.9-1.0 lag fraction from the real-world signals we are likely to generate. There is a zone of good performance at about an EM length of 20-25, but it is narrow, and it falls off swiftly.
This zone that we propose to trade is in the right forefront in the 3D surface plot. Two issues are immediately obvious. The tradescape uses what we can regard as an ideal signaler in terms of mining the order within the time series. A real-world signal is not likely to be as accurate, meaning it is probable that it will perform more poorly. Already we are perched upon a somewhat narrow feature of the trading response surface. We would like the zone we target for the signaling system to be a good deal more robust, which means wider, more of a plateau and less of a ridge. Secondly, there is the more obvious point that we would love to see the reward to pain a good deal higher, if at all possible.
Trading an entity that represents a country's markets would not be expected to exhibit significant asymmetry. This is one advantage to trading an entity that represents an average of many entities. One can usually use a symmetric signaler, one where the same information content is used for both entries and exits. In this grid of asymmetries, the center panel of the 25 has the unit asymmetry. Since the entity that will be traded, EWA, favors signal symmetry in terms of time horizons for entries and exits, we will use Asymmetry=1 from this point forward.
The Referential Observation
For this example, we will hypothesize that a better or more stable signaling target might be the EWH, the Hong-Kong ETF. A good portion of Australia's economy is based on supporting the growth occurring in Asia with its materials sector. Hong Kong has also been closely connected with Australia from British colonial days, so the connections are sound along political, economic, and fundamental lines, and with the change in Hong Kong's status, it is even now more closely mirroring China's emerging markets.
We thus take the first step of confirming a connection by generating a referential tradescape. We trade EWA using EWH as the surrogate, as the target used to generate the actual signals.
The differences may seem small, but there are two observations worth noting. First, EWH has slightly better function about 0.9-1.0 lag fraction. Surrogate signaling is not only viable here but it is an actual improvement despite the "fuzziness" it typically adds to the signaling process.
Viewing the 3D surface plot helps to illustrate that the narrow ridge in the far right foreground, at lag 1.2 and EM length 20, is not quite as narrow. The signaling response surface is a bit wider, though far from ideal.
It is worth noting that referential signaling, where a surrogate is used instead of the traded entity directly, is typically done on individual securities rather than this type of ETF. The surrogate, usually some form of index or entity representing an overall national market or market sector, is used to remove the influence of chaotic movements arising from news or high trading activity specific to the individual entity.
The Sentiment Observation
For this exercise, we will explore whether the surrogate (EWH) is also suitable as a sentiment target.
In our experience it tends to be a rather costly choice to disregard sentiment in any trading system. What traders decades ago called trend filters, we put in a more general classification of sentiment signaling, the partitioning of time into zones favorable for long trades and into zones favorable for short trades or being out of market.
A sentimentscape studies the value of a two-stage signaling system where this kind of partitioning is implemented with the same lag fraction as the primary signaler (the signaling occurs on two very different time-horizons, but they share the same fraction of lag relative to their respective EM lengths). The coarser time horizon signal partitions time into windows where only long or only short trades are allowed. The finer time horizon signal generates the entry and exit signals for the trades within those windows.
If no sentiment length is specified, a sentimentscape consists of an array of predefined EM lengths for the sentiment signaling. Looking closely, we see two key items. A very fast sentiment filter (EM length=25, second panel) offers significantly greater reward-pain at EM length 10-15, but a lag fraction of 0.8 or better is likely needed. We also note that the EM lengths 50 and 65, especially length 65, produce a very robust surface.
This is the sentimentscape for trading EWA with an EM=65 length EWH sentiment filter.
The probability partitioning has somewhat improved the reward to pain shown in the RRt surface. In a sentimentscape, the sentiment filtration is applied only at lengths below the length of the primary signaler. There is thus no sentiment filtration on EM lengths of 65 and higher. At these higher lengths, only the primary signaling occurs.
Do We Have a Winner?
Simply because a tradescape looks robust, doesn't necessarily mean that such is the case. We can say the average behavior across a wide span of time is robust, and that is positive. Using this Asian market as the surrogate and sentiment source, we would have traded for more reward than pain at virtually every signaler setting except the very fastest, and this would be true for even very weak signaler.
What we don't know is how this performance distributed across time. We could have seen all of the benefit in just the first half of this ten year period, or it could have wildly varied. It is hard to stay with a signaler that is not a consistent winner.
progressive tradescapes offer the means to study tradescapes sequentially across time.
This is the progressive tradescape for directly signaling EWA, no surrogate, no sentiment filter. Clearly the Australian market, or at least the entities representing it within EWA, have changed behavior dramatically across the last 10 years. In terms of trading for more reward than pain for these four 2.5 year periods, we see the latest as particularly grim.
This is the referential progressive tradescape where EWA is referentially traded using EWH. Here, in general, we see an improvement from the surrogate.
This is the referential progressive tradescape that includes both the surrogate signaling and the sentiment filter. We clearly fare far better in the most recent 2.5 year period, and we also see some promise of trading for more reward than pain for each of the four 2.5 year periods. The EM length of 20-25 is very good in three of the four periods and there is at least more reward than pain in that zone in the second panel.
And so, do we have a winner? That all depends on whether or not one can construct a real world signaler and a real-world sentiment filter with these time horizon and lag properties. We do know there is promise substituting an Asian market index for both the primary signaling and for generating sentiment windows.
At this point a few notes are in order. What was accomplished by this analysis? What must now be done?
First, a tradescape is a picture of what can be done with the orderly trending in a time series. For EWA, we saw that it was a difficult entity to directly signal, at least if we set a requirement that we seek to trade for more reward than pain. We saw that trading EWA was not a simple problem.
We accomplished a good deal in the set of analyses. As a result, we know that the reward to pain and the lag tolerance can be improved by the combination of using EWH as a surrogate and also using EWH as a sentiment target. We know that such a favorable place exists historically, even if we haven't yet built the real-world system that will exploit it. In many ways, therein rests the value of tradescape technology. Before we invest any time in designing the real-world signaling system, we know that we may well not have to jump through hoops to develop a leading edge signaling algorithm. By the proper choice of signal target, and with the proper choice for adding sentiment probability, a simpler signal algorithm might very well suffice.
We also know that the nature of the time series is such that the range of optimum time horizons for the primary signal is somewhat narrow. Although we found a way to expand that somewhat, we begin with the understanding that any system will need to be tuned. This is the kind of entity where picking a popular breakout level for an entry and exit and running with it would be foolish. The robustness simply is not there.
We did not look at other surrogates in this example. There may be other Asian markets that work more effectively. Although perhaps not as close a fit fundamentally, the Malaysian and Indonesian markets move with a higher measure or order and are thus easier to signal. It takes but a few seconds to know.
We did not look at other sentiment targets in this example. There may be better market entities to use than EWH. A US market entity might be worth exploring.
Once the surrogate and the sentiment target are selected, the design work can begin. We work outward in. We know we want an EM length 65 sentiment filter that uses EWH. We can generate an EWH tradescape, find the EM length 65 properties, and use the signal evaluation procedure to test the various sentiment signal algorithms at settings close to this time horizon. The average trading length is about 2.5-3x the EM length, meaning we want a sentiment signal whose windows average about 175 trading days. Once this filter is designed for EWH, we can move to the finer time horizon where the primary signaling algorithms, also operating on EWH, can then be tested on EWA. We know in advance that we want something at about an EM length of 20-25 days.
Finally, note that the referential sentimentscape represents an ideal in terms of accuracy for both the primary and sentiment signals. Real world signals will not exactly match the turns in the primary and sentiment signals used in the sentimentscape. An error of just 10% in terms of missing turns equates to a potential 20% error in terms of overall signaling (both the entry and exit have to be right to effectively book a trend). For those trend-following traders that want to trade the order within a time series, we feel that the tradescape technologies give a good picture of what a realistic upper limit is in terms of performance.
While it is possible to find real world signals that outperform the tradescape surface at a particular set of parameters, in general any given real-world signal algorithm mapped to the same time horizon and lag scaling, will typically fare more poorly. That is expected. The tradescape technology adds lag by shifting ideal signals in time. The density of the lag at the turns is thus an impulse function at every point in the tradescape. There is no scatter. The lag in the signal plotted at any given point in the tradescape will be identical at every turn or signal transition. It is not an average lag, but an exact one at every turn. We thus see what a uniform lag generates for performance of a signal that will have as much accuracy as the order within the system allows. For an unlagged signal, that is typically a 90-100% win rate.
In the real world, except for the simplest MA algorithms, the lag at the turns within a signal will have a defined scatter or density. How the lag is distributed at the turns can make a very significant difference in performance. Adaptive algorithms can sometimes shift the density of the lag at the turns in a favorable direction, but it is an art to have that happen consistently and robustly. Certain adaptive variables, such as volatility, fractal dimension, the acceleration of momentum, buy/sell pressure, and so forth work well in certain market states and not others, and the net effect across time may well be a wash in terms of this density or scatter of lag.
There is another factor, however, and this is the inaccuracy that arises from the adaptation being in the wrong direction. In such instances, it is not merely a matter of having to live with a slower entry or exit. The transition in the signal may not happen at all. The ideal EM signal maps each transition and then shifts it in time for the lags shown in the surface. The transition is thus present at every lag. In other words, accuracy is maintained at all lag fractions. The ideal signal is simply shifted or delayed in time to simulate this uniform lag.
The main problem with an adaptive signaling algorithm is that it loses accuracy as it seeks to be more responsive at certain points in the time series. That can prompt premature exits that lose portions of an existing trend, or premature entries that jump in before a trend has fully formed, at least in its probabilistic sense.
This is why the signal evaluation routine in the tradescape package goes to such lengths to report the accuracy at the turns in the real world signal. If the accuracy is low, but the return is very good, this helps one to know that it is not the order in the movements being successfully traded, but rather the chaos or less orderly movements. That suggests a very strong need to eliminate the possibility of random chance producing the positive result.