Platform Signal Design
TradeStation Analysis Techniques
Separating Order from Chaos
Most trading signaling systems seek to trade order and defend against chaos. In a mathematical sense, order is defined as a central deterministic channel which is smooth within a sea of seemingly random or stochastic noise which represents the actual prices. The smooth central channel is deemed to be the order, the 'signal' in the SP sciences, and everything else is the chaos, the 'noise' to signal scientists.
When trading order, one determines the direction of that central channel and jumps on and off as appropriate. The order is the human driven underlying trending that is present in the time series.
The idea behind using EM (expectation modeling) technology to create tradescapes is that it is possible to answer most of trading design issues in an idealized system, one that seeks to perfectly trade only the order within the price movements in the time series. By using such a system, it is possible to map the entire trading landscape for a given financial time series, provided one knows what is important and how to properly carry out that mapping process.
Many trading system studies are carried out with a set of trusted or as yet unexplored real-world signaling algorithms and the outcome is often either weak or unsatisfactory performance for all signals, or the suspicious condition where just one or perhaps several show sufficient promise. At this stage, the signal designer has usually invested a large measure of computational work, and may have learned relatively little.
The reason is that the time horizons in the signal series might have been all wrong for the entity being traded. The bar density might have been unfavorable. The issue of accuracy versus lag is intrinsically and fundamentally at the core of the design process, but in general neither is known. There is no gold standard by which to draw a comparison. In general there is no reference at all by which a comparison can be made on the signal level to know if one actually did well, given the nature of the entity to be traded, or if one fared poorly.
Traditionally, one uses backtests to improve upon a system that is already in use actively trading a given entity. That existing system becomes the reference and any comparison is generally only by backtests and the bootstrap type shuffling of blocks of data. Overfitting and ill-posed optimizations are serious issues if one finds a system rising to the top of the pack as a result of the fortuitous processing of fat-tail events.
In most cases, a trading system is rendered active without knowing how efficient it is in terms of using the information included in a sliding time window. Clever algorithms adapt their time windows for specific conditions hoping to diminish lag at appropriate times and to better tolerate it at others. Most algorithmic trading systems go on line with no foreknowledge of the signaling lag that actually occurs at the turns where orders are executed and even less is there any appreciation for the extent to which those turns are historically accurate. Whipsaws are an example of a real-world phenomenon that can dramatically impact the accuracy of any signaler.
Even in the instance of long-history optimizations where one might map all possible lengths for the two components in a moving average crossover, or the window lengths for entry and exit breakouts, the 3D response surface will map only a small region of the landscape, and each point mapped will vary in both accuracy and lag, neither of which is generally estimated.
Mapping the Ordered Trending in a Time Series
A tradescape is a 3D picture of the trading landscape for an idealized signaler at various time horizons (information content) and introduced signal lag (enforced inefficiency or latency). If an algorithm can be constructed to generate close to complete accuracy in terms of catching nearly all of the turns in the price movements, it is possible for that map to represent the true order in the trading at every length of time window and at increasing levels of signaling inefficiency.
Even though a ideal EM signal is not a real-world trading signal, it's lagged signals can be viewed as a gold standard for any real-world signal. A set of well-crafted EM signals define the expected performance in terms of the time horizon best traded, and how good one's signaler must be in terms of lag to potentially realize a desired reward-pain.
In essence, a tradescape tells you what can be realized from trading the order within a time series. It tells you how good you have to be with your signal design, assuming you want to capture all of the order and none of the chaos in the price movements. It is a benchmark, a target. It can be beaten by signaling systems that trade both chaos and order in that very effective dance, but in our experience this is rare. A tradescape represents a respectable upper limit of what is achievable with any trading system signaler.
Trading Sciences Technologies has more than five years worth of development work invested in the Tradescape platform. It took several years to build an algorithm without whipsaws, and to do so without the epsilon, confidence, or volatility band signaling traditionally used to defeat the undesirable signal cycling. Such bands mute the chaos that slips into the expectation and is inadvertently modeled.
The Value of a Full Trading Landscape Map
An ideal signaler maps the entire trading space from the theoretical and wholly unachievable (there are no zero lag ordered signalers) to the zones where only signalers with terrible lag are found.
That entire domain is covered by an EM signaling algorithm that is absolutely fixed. There is no optimization; the adaptive properties of the algorithm are general, not specific to the entity being signaled. The map of the trading landscape is as close to ideal in terms of accuracy as possible in terms of detecting the ordered turns in the price movements. Even with zero lag, 100% wins do not and should not be expected to universally occur (there have to be instances where chaos supersedes all order even in a theoretical signaler with zero lag), but in general very high win rates are observed at very low lags, and win rates approaching random chance are seen at very high lags.