The Practical Application of Intelligent/Predictive Agents

[UrmaBlume 10]

The same intelligent/predictive technologies that track missiles in 3 dimensions at mach 2 can be used to look at least a few minutes into the future of the price/time/volume continuum that describes today's world financial markets.

 

The chart below demonstrates the implementation of some of these technologies in an application we use to teach and trade what we call "Scalping with an Edge."

 

The charts and text below will describe the process required to build such an application as well as some of the tools used in its construction.

 

As an example of the application of intelligent agents to processed market data the shot below is of 30 minutes' trade in the ES on 3/18/2010. Times are PST.

 

The bars are 2k volume bars. Each bar represents about one thousandth of average session volume. This makes them about 24 second bars.

 

The color of the bar represents the dominant bias in that particular time/volume frame. In the chart below you will see gray bars (no dominant bias), red bars (sell bias) and blue bars (buy bias).

 

We define TradePoints as those points in the price/time/volume continuum where a specific action is recommended.

 

In this case the blue and red + is a recommended entry price. The small red dot is the recommended stop and the blue a good point to scalp. The text defines those points for the current bar. Both the text and the TradePoint dots are automatically posted at the first tick of the bar and updated every tick. When there are no recommendations, as it is with much of most days, the auto text posts "No Trade" as is the case in the second chart below.

 

The middle sub-graph shows a calculation of net new trade by commercials over a multi-session time frame.

 

The bottom sub-graph shows a collection of weighted biases from 7 different time frames. The yellow is cautionary, the red - sell bias and the blue - buy bias.

 

The percentages to the right of the current bar indicate how close the bar is to completion as well and the percentages of both buying and selling volumes within that particular bar so far.

 

 

In the graph below we have just come off of a Buy signal and the recommendation of "NO TRADE" is posted. Note the bars are gray - no particular bias and the absence

of TradPoints indicates a lack of power to follow through.

 

 

Data Pre-Processing

 

This is the most important of all. Whether the agent is a neural network or regression machine it does much better with well prepared inputs. For the development of these models outliers in the "seen" or training data are removed. Also the input data is scaled to resemble the target data. This is where the output of some fairly traditional algorithms are normalized to match target data. An example of this is time of day normalization of the of average volume into percentages as is done on our Market Heads-Up Display (HUD) and also described in this thread here on TL

 

 

Modeling Process

 

Today's Quants have a range of tools and algorithms available to them to aid in the development of their models. One site that can be very helpful when selecting or trying to evaluate these tools is the KDNuggets site. There are neural networks, Bayesian Networks, Genetic Algorithms, Rough Sets and

routines that implement Fuzzy Logic. Generally we use MARS (Multivariate Adaptive Regression Splines) from Salford Systems in San Diego. It has scored very well in the annual competition at the KDNuggets site, is very easy to use and we have built a script/application that automatically converts is ouput function to Easy Language so that with a cut and paste the models can begin to function on line and make their predictions in real time.

 

 

Rules Generation

 

The output from the various models is sent to one or a series of rules generation technologies. We use genetic algorithms, decision trees, CART from Salford Systems and a rules generator with the improbable name of WizWhy from WizSoft to optimize and to generate the rules that generate the trade points shown in the charts above.

 

The chart below shows trade during the first hour of Friday, 3/19/2010. While this chart shows better than average

results, it does indeed also show the kind of accuracy that is possible with the help of well constructed intelligent models and other predictive technologies.

 

 

 

Good Luck

 

Cheers

 

UrmaBlume

Edited March 20, 2010 by UrmaBlume

[UrmaBlume 10]

Here are Today's TradePoints:

 

The + indicates entry price

The small blue dot is take profit

The small red dot is the stop

 

 

[Tradewinds 12]

The same intelligent/predictive technologies that track missiles in 3 dimensions at mach 2 can be used to look at least a few minutes into the future of the price/time/volume continuum that describes today's world financial markets.

 

UrmaBlume

 

Are you using something like the weighted least squares approach for best fitting a set of points to a non-linear approximation? Missiles don't have a choppy, erratic trajectory. Price points in the market can either be uncorrelated and uncertain, or have some reasonable distribution. If it is determined that those price points are a candidate for a normal distribution, they can then be fitted to a predictive curve.

[Tradewinds 12]

Generally we use MARS (Multivariate Adaptive Regression Splines) from Salford Systems in San Diego. It has scored very well in the annual competition at the KDNuggets site, is very easy to use and we have built a script/application that automatically converts is ouput function to Easy Language so that with a cut and paste the models can begin to function on line and make their predictions in real time.

 

Looks interesting. I guess I should have read further.