About Tradescapes

The Tradescape concept took five years from conception to software. Although we finally have technology that addresses the critical questions that we always believed could be answered, the path hardly followed the rational track we might have expected when we began.

In every aspect of data analysis and statistics, fractal science is everywhere: fractal scales, self-similarity, information theory. There were the key issues with any form of information, accuracy and latency (this elusive measurement of lag). We felt there had to be a sane and effective way to answer the question:

“What is the robust reward-pain we can realistically expect at the lag of our real-world signaling systems when trading a given entity at a given time horizon?”

We knew we would have to build an ideal or at least close-to-optimum theoretical signaler which would define both information content and lag, and it would need to separate the chaotic movements from the ordered ones in a way that would offer the highest overall trend-based return across all financial instruments. If we could identify such an algorithm, we could map the tradable order or trend-favorable trading landscape of any entity.

We ended up devising our own algorithm. It is authored in C++ and consists of nearly ten thousand lines of code for just the basic numeric modeling. The purpose of this section is to introduce this science, and hopefully offer a few surprises, trusting that you will be surprised as much as we were in the course of this evolution of Tradescapes.

Let’s Do a Signaling Optimization

We will start with what is now being more routinely done, a two-dimensional parameter optimization for a signaling system. We will use QQQ and 10 years of daily data. We will look only at the HH and LL breakouts in a turtle-type trading system:

We have a basic 3D response surface for long trading of QQQ with the HH (entry) settings on the x-axis and the LL (exit) settings on the y-axis. The surface is made up of the outcome of more than 3000 different backtests. For the z-axis reward-pain, we use a robust annual trend or CAGR as the reward divided by the average retracement as the pain. In this 3D view, we look down upon the response surface and we readily find the typical optimization (and “overfitting”) issues. Those are clearly narrow peaks and ridges!

Let’s look at a portion of this optimization in a patch plot. We can try to take a call on where robustness is best found. If we have any measure of insight, we know we have as much chance of seeing a map of how the optimal parameters have shifted or moved about across time as we have of seeing multiple sweet spots that have been constant across time. In ten years history, we don’t expect to see a strong performance zone from random chance. We believe there was legitimate trading function at the different favored response surface zones.

What We Know and What We Don’t Know

We know that there are three favorable HH,LL parameter regions in this ten-year history. The one that works best in the most recent history might be the obvious choice. Let us say, for example, that we like [11,40] for its potential robustness across this extended historical period. We do not pick the parameters of highest reward-pain, but rather a zone we feel is likely to remain profitable with the inevitable shifts in behavior across time that we would expect to occur. We try to anticipate and accommodate those shifts.

We also know that breakouts use the information content in a very simplistic way. We would expect a wide scatter in the lag since the upper and lower bands at any point can arise from any point in the two sliding windows. Still, we know nothing about the lag that we actually realize with an [11,40] parameter pair. We know that breakouts trade a mix of orderly and chaotic movements.

We have no idea how much of the trending or order in the time series we capture with the algorithm. We have no idea how accurately we signal the order that is traded. And those are precisely the questions we felt we should be asking.

A Leap

At this point, let’s assume that we have the algorithm that allows us to answer those questions. Let’s revisit the optimization in terms of the two parameters that we feel ultimately matter most. We want to know the information or time horizon of the signaling system in an absolute sense, not something inferred from the average trade length or that which varies with whatever algorithm we happen to be using. We also want to know the average lag we achieved as well as the scatter in those lags, and we would prefer robust measures not subject to being skewed by outliers.

Here we transform the [HH, LL] coordinate system to an information content (time horizon, the EM length) and lag fraction (lag as a fraction of this time horizon). All three thousand plus points are plotted here. Clearly, much of the time horizon – lag landscape failed to be represented by the large matrix of breakouts tested. Further, we see that the return was hardly such that a useful interpolated response surface was possible. Even when the lag is quantified, each of the pairs will generate a very different accuracy in terms of trading the order or trending within the prices. Two points can have the same lag and the same information content, but have widely different reward-pain because of this issue of accuracy in signaling the turns.

Let’s Try Harder

We would love to cover the whole of the time-horizon and lag trading landscape with the real-world breakouts. As such, we increase the maximum breakouts lengths in the study and we now have 6500 backtests, each with a specific breakout pair. We further color the points by delta RRt, the reward-to-pain improvement over the underlying buy and hold for the ten year period. For this plot, the gray points produce an inferior reward-to-pain. The blue points produce anywhere from 0-0.5 improvement. The yellow points produce a 0.5-1.0 improvement. The red points produce a 1.0-1.5 improvement. Red is good.

Lower lag does matter. The lowest we achieve for lag is 60% of the reference EM length. There is no near-zero lag. The pattern is odd, but we see EM lengths that represent favorable time horizons and those that are much weaker. We can also look at each of the points and note the accuracy of each signal in terms of catching the order within the trending.

The important point is that we now have an entirely new paradigm for robustness in a trading signal. We choose the signal offering a favorable time horizon and lag, and which generates a good return in that zone.

It took the better part of a day to generate this plot.

The Tradescape

This plot required about 1 sec to generate from TradeStation. Every point represents as close to full accuracy as we can realize with our EM algorithm science. The scale is the same as the transformed breakout optimization. Note that a true EM-based trading landscape is continuous and smooth. The signal is theoretical insofar as its purpose is to map all tradable order in the time series to give a complete picture of the well-behaved trending component within the data. While we cannot trade this idealized universal EM signal, we can see what our entity looks like if we embark on the task of building a signaler to trade the ordered trending in its price movements.

Note that the odd pattern in the breakout optimization largely follows the perimeter of exceptionally high reward-pain in the tradescape. One is not going to enter this rarefied region with a basic signaler. A remarkable one is required. Still, we see what can be expected depending on how good one’s signaler happens to be in terms of lag. We see where one wants to work in terms of information content. We see two clearly defined zones just as we found two principal regions in the original breakout optimization.

A Paradigm Shift

We now do something entirely new. We select an EM-based information content to trade. It may be an EM length 7 or 14 to seek to go after the finer reward-pain in the exceptional area. Or it may be the more conservative EM length 9-11 that captures the fuzzier peak at a higher lag fraction.

Life is now simpler insofar as we have identified a time horizon based on a universal signaling algorithm. We have taken accuracy out of the process as we begin. We assume we are looking at what we would achieve if our signaler were 100% accurate in catching the turns in prices at each time horizon.

We still have the work of identifying the real-world signalers that can offer the necessary time horizon, and with hopefully a lag and accuracy that allows a good measure of the benefit to be realized. In practice, however, a given class of signaler with a given set of parameters will generate close to the same EM time horizon across a wide variety of entities. One soon learns what can be expected from one’s favored signaling algorithms and strategies at their various settings.

Here we plot the signal analysis from eight different breakout pairs. The standard turtle [20,20] and [55,20] are signals 1 and 2. Signals 6, 7, and 8 are the [10,20], [17,40], and [11,40] we selected by looking for robust zones in the original breakout optimization. When plotted atop a tradescape and colored in accord with the tradescape gradient, we see these three signals make sense. They don’t achieve what is possible from trading the order in the trending movements, but they do well: there is more reward than pain. Note the grayscale points are profitable, but have more pain than reward.

The idea is to use a universal EM signaler and its tradescape as a far simpler first step. In a few seconds, you will know your choice(s) for a time horizon and how good your signaler will have to be to even have a shot at giving you the reward to pain you seek.

You may discover there is insufficient order or trending (too little lag tolerance), or too little performance (too little reward-pain), to bother further, given what one soon learns to be the limits of one’s signaling algorithms. In a few seconds instead of hours or days, you can move on.

If you like what you see, you can then analyze the intended trading signal atop the tradescape to insure that one is in the intended place for time horizon and lag. There is no optimization or fitting in the typical sense, only choosing the place upon the trading landscape where you wish to reside. With some experience, it is possible to have a good confidence going forward.

However signalers seek to address chaotic movements, many classes of signalers fundamentally trade the order within the price trends. These include all direct price signalers, most price envelope signalers, such as breakouts, and most sentiment signalers, such as MA crossovers. With experience, hours or days of signal design can be reduced to minutes, and with a higher confidence.

There is the common issue where the lack of accuracy results in less than the potential that can be realized at a given time horizon and lag. With an EM reference, signal analysis and statistical techniques can be used to ascertain not only the lag, but also the accuracy in terms of catching the turns. One can then work on algorithm refinements to improve the accuracy, or perhaps reduce the lag. Soon enough you will discover that the Heisenberg principle is alive and well in the fuzzy accuracy-lag tradeoff.

The EM tradescape science can be used to study signal asymmetry, as observed in this breakout example. The long-term optimum for information content asymmetry can be discovered in a succession of tradescapes that generate signals with different ratios of entry and exit EM lengths.

More importantly, tradescapes can be used to study an entity’s behavior across time. One need not merely target the long term historical sweet spot. One can look for a good place in all segments across time.

Intraday tradescapes can be used to determine the best bar density to use to target a given time horizon. Analyzes that would typically takes many hours or days can be done in minutes, and with strong confidence in the outcome.

Given that each point on a tradescape surface, and each real-world signal analyzed in conjunction with a tradescape, is an actual backtest, it is a simple matter to explore any selected point on the tradescape by viewing the equity curve which generated it. This ensures one is reasonably pleased with the more recent performance in time, and to better accommodate blind walkforward studies.

Summary

In this basic introduction to tradescapes, we discuss a paradigm shift where the parameters of signalers are replaced with an absolute information content and lag fraction and we further show that it is possible to realize this map or landscape of the tradable order using a universal signaling algorithm from the EM science. We introduce the concept that what is important in designing a trading system is the time horizon of the signaling (this includes the density of the bars in the data sampling), the lag fraction of that time horizon realized by the signaler, and the accuracy of catching the ordered turns with that algorithm.