Tradescape Platform Signal Design
TradeStation Analysis Techniques
The Tradescape Paradigm for Signal Designers
The Tradescape Platform Paradigm is one that addresses many aspects of the trading science. Here we specifically discuss signal design using tradescapes.
A signal design can be skewed by coincidence or random luck. For example, the turn in a signal can result in a strong win even when it does nothing more than originate from an undesired whipsaw. That entry may just happen to coincide with a fat-tail event in the direction of one's trade. A careful designer worries about this random influence. Especially when a design includes a small number of trades, a little coincidence can go a long way.
Further, optimization is often employed in signal design. One fully expects the signal derived from an optimization to not fare as well walking blindly forward. A great deal depends on how well the optimization catches the long term aggregate behavior of the entity and trading signaling algorithm. If the optimization responds to any fortuitous combination of signaler parameters and fat tail events in the entity, these chaotic events can seriously skew the optimal signal settings.
In a tradescape signal design, we avoid the burdensome process of conducting a full signaling algorithm parameter optimization. We know such an optimization can respond to fat tail events and skew the signal parameter settings accordingly. Many response surface optimizations generate plots with multiple "hot zones" and "cliffs". If you have worked with such optimization surfaces, you have likely found yourself seeking in rather an artistic way to locate zones that were robust. You would try to center your parameters in such zones, well away from the sharp cliffs, even if they were also removed from the actual peaks. You would try to position your signal parameters so that the optimum could shift around a good deal and the performance would still be good.
In a tradescape signal design, we avoid the burdensome process of using Monte Carlo design methods where the analysis (in-sample) and reserved (out-of-sample) periods are endlessly swapped around. If you ran a thousand MC simulations, achieved a thousand different parameters from optimizing each of those different analysis periods, and then tested those in a thousand different reserved periods, you would likely have accomplished a different way of achieving what we achieved by the art of choosing a robust zone on a single optimization map. Even with such an MC approach, there can be uncertainties. When there are few trades, as in scenario using EOD (end of day) data, it is possible there will be a wide scatter in each of the design parameters across the different experiments. That does not lend confidence. There could be a sense any performance was more random than real.
The full parameter optimization and the MC-based design are viable approaches and we respect them. We've long used them. The problem we saw with both of the approaches is that a huge effort (or at least a great deal of computation) went into arriving at a signal. In some cases, the whole optimization showed nothing of merit. You could potentially spend a great deal of time testing a new strategy against a large pool of candidate entities, and walk away with nothing even promising.
A New Signal Design Paradigm
The new paradigm for tradescapes begins with an ideal signaling algorithm designed to separate the chaotic aspects of the price movements from the ordered behavior. We use an algorithm that we designed to respond only to the order in the time series. We crafted such an algorithm to give as close to full accuracy as possible across all instruments and markets. We then used this EM (expectation modeling) reference algorithm to map the entirety of the practical trading landscape. We created tradescapes.
We did so with two specific specializations. We wanted the map of the trading landscape to consist of full accuracy in terms of catching the order in the price movements. Further, we wanted this EM signaling algorithm to full exclude the chaos within the movements.
In terms of moments, we determined a maximum likelihood method for estimating the ordered portion of the skew or third moment and to retain that portion of the trend and disregard the remainder. Similarly, we determined a way to estimate the kurtosis or fourth moment and to retain that portion of the fat tail events most likely to persist. For tradescapes to work, we needed the signaling to consist of a response that was limited to only the order in the price movements.
Our definition of order corresponds with trend-following. You enter, with some measure of lag, once a sufficient movement in the direction of trade occurs. You exit, again with some measure of lag, when a sufficient movement against the direction of the trade takes place.
Using this approach, we generate a robust trading landscape that clearly shows the impact of time horizon (information content, the data visible to the signaler at any point in time) and lag (the latency or delay in the responsiveness of that signaling) on the trading performance.
Each tradescape maps about 700 EM signal backtests, each with a specific time horizon and lag, to a performance metric, usually a reward-to-pain ratio.
Using tradescapes, we dramatically simplify the design process. It takes about a second to generate a tradescape. In that trading map there will generally be one or more sweet spots if there is sufficient order in the price movements to be traded for order. A tradescape can immediately show you which entities have not shown good potential for behaving orderly in terms of successfully entering and booking trades using the bar density in the input data. It will tend to show that almost all financial entities can trade order very effectively at low lag. It is when the lag is higher that the "fickleness" of the entity's behavior becomes apparent. We call an entity that can trade order with a basic signaler "lag tolerant". If that property is present at all, it is likely to be so only at a sweet spot, a time horizon that responds favorably to trading order. It is possible to sometimes find that time horizon with daily data. For certain entities, it may only exist on the faster time horizons accessed by faster sampling or intraday bar densities.
The Steps to Tradescape Signal Design
In the design paradigm using tradescapes, the first step is to determine if an entity can be traded for more reward than pain at the effective lag fraction you expect to see with your real-world signaler. These dimensions of time horizon and latency are fundamental ones. You will quickly come to know what you can expect from your real-world signaling algorithms and strategies in terms of lag and accuracy. For a given system that you have come to know, you can know in a few moments whether it will be successful in a consistent and robust way with the candidate entity you are exploring. One look at the entity's tradescape will tell you if you are likely to be successful.
Once you have identified the target you will trade, you spend a few moments determining the time horizon to trade. This is as simple as visually identifying the sweet spot in the tradescape. If there are multiple ones, you select the one you can realistically achieve with your signaler. That is the primary design aspect to tradescapes. You simply identify the time horizon you want to trade.
It is related to principle cycle analysis, except you are identifying the information content or time horizon the signaling algorithm must have instead of a fundamental sinusoidal-based wavelength.
The signal evaluation routines in the tradescape package take your real-world signal and determine its time horizon and lag with respect to the tradescape signaling algorithm. It places the real atop the theoretical. You know what your trading signal 'uses' in terms of its information content, and it is measured using this absolute reference of the tradescape time horizon. This EM length time horizon universally means the same thing, across all signaling algorithms, across all entities traded. That in itself represents an enormous leap in this science.
You see how 'responsive' your trading signal is in terms of its average lag at the turns where entries and exits actually originate. You see the real world performance as compared to this theoretical system that trades just the order at this same time horizon and lag. If your signal effectively trades both order and chaos, you may see the tradescape performance exceeded by your signal's trading performance. In the real world, you will have some measure of inaccuracy in catching this order. Further, the lag will not be uniform but will have a scatter associated with it. The signal analysis routines furnish you an estimate of both lag and accuracy. You can tweak design systems to improve upon either or both. If the enhancement is sufficiently universal, it should show itself across the board with all entities traded.
When you explore the tutorials in the tradescape package, you will see this signal design paradigm in action.