Using ARM3 to Improve NSP-31

Introduction

This Technical Article illustrates how to we improved the NSP-31 strategy with neural network techniques.  First, we discuss how the strategies were configured and trained, including the introduction of new features for the NN Score block in OmniTrader.  Then we present equity simulation results for each strategy.

Configuring the Strategies

Three strategies were considered.

·         The base NSP-31 strategy is comprised of an NSP block for signal generation and an Orders block to generate exits.

Base NSP-31 strategy.

·         The NSP-31 NN v2 strategy includes an NN Score block for signal scoring and filtering.  Note that the NN Score block can now also be placed at the end of a strategy flow chart.  In this case the block will act as a filter on entire trades (i.e. signals and their orders).

NSP-31 NN v2 strategy: a neural network scores and filters signals along with their orders.

·         In the NSP-31 NN Stops strategy, an additional NN Score block - separately trained on a bar-by-bar basis – is placed upstream from the Orders block.  Because the purpose of this second neural network is purely to assign scores to each bar in the vote line, the internal score cutoff value is set to zero.

NSP-31 NN Stops strategy: an additional neural network assigns scores to every bar.  The score is the used to adjust stop levels by stops in the Orders block.

Within the Orders block the NN NSP Trailing Profit Stop is enabled to replace the NSP Trailing Profit Stop used in the previous strategies.

 Orders block showing the enabled NN NSP Trailing Profit Stop and its parameters.

This new stop works in a similar fashion as the NSP Trailing Profit stop, with the only difference being that the cushion value is adjusted at each bar, based on its advisor.  In the NSP-31 NN Stops strategy above, the advisor – or score – is assigned by the NN Score block that precedes the Orders block.  The Mid Cushion and Max Cushion parameters specify the trailing profit stop’s cushions at scores of 50 and 100, respectively.  At any bar within a trade the cushion is recalculated via a linear function within the specified points.  Negative cushion values are not permitted.

Note that identical exits can be established in Trade Plans.

Performance Summary and Equity Simulation Results

The following table shows the performance summary for the three strategies which includes all trades between January 2001 and December 2006.  The original NSP-31 test and validation list was employed (S&P100 as of September 2006).

Performance Summary for the three NSP strategies.

Equity simulations were conducted over the same test period and symbol list with 100% of Equity allocation method.  Screenshots of the resulting equity curves and statistics are shown below.

Equity simulation results for NSP-31 strategy (see Summary #2, below)

Equity simulation results for NSP-31 NN v2 strategy (see Summary #2, below).

Equity simulation results for NSP-31 NN Stops strategy (see Summary #2 Below)

Performance Summary #2

The relevant statistics from the above screenshots are summarized in the following table.

Portfolio simulation statistics summary for the three NSP-31 strategies.

NT = Number of Trades
HR = Hit Rate (Accuracy)
ROI% = Return on Investment%
WDD% = Worst Drawdown Over Entire Period
AvgInv% = Average % of Equity Invested (200% is the maximum)
 

Validation of Neural Network Predictions

The neural networks for the described strategies were both trained with data from August 2002 to July 2004.  Though some of the improvements were due to the networks’ prior knowledge of the system’s performance during this time span, it is evident from the previous screenshots that the employment of NN Score blocks is beneficial throughout the test period, resulting in considerably higher returns as well as lower drawdowns.  The following scatter plot shows the relationship between the neural network’s predicted score of the NSP-31 NN v2 strategy and the recorded profit for all out-of-sample trades (i.e. signals not used for training the NN Score block).

Profit vs. Score scatter plot for out-of-sample data collected for the NSP-31 NN v2 strategy.

Fitting a trendline through the data points shows a positive correlation between the score assigned to out-of-sample signals and their respective profit.

NN Score Allocation Method

Well-predictive networks can also be used in the design of the allocation method employed to trade a particular strategy.  As an example, consider a method that allocates a percentage of available equity to each trade, where the percentage is chosen based on the signal’s score.

In the following equity curve screenshot and statistics table, we compare the results from simulations obtained by trading the NSP-31 NN v2 strategy with a fixed 50% of equity and by trading the same strategy with a percentage of equity that varies linearly between 0% and 100% based on each signal’s score.

NSP-31 NN v2 equity simulations for fixed and NN Sore-based % of Equity allocation methods.

Portfolio simulation statistics summary for fixed and NN Sore-based % of Equity allocation methods applied to the NSP-31 NN v2 strategy.

The same method can be applied to the NSP-31 NN Stops strategy (see table below)

NSP-31 NN Stops equity simulations for fixed and NN Sore-based % of Equity allocation methods.

Portfolio simulation statistics summary for fixed and NN Sore-based % of Equity allocation methods applied to the NSP-31 NN Stops strategy.

Other uses of NN Score for Exit Management

The present document has shown a way to improve exit management through the use of the NN Score block.  While the example and results employed a custom stop coded via the SDK in VB.NET, the advisor is also available for coding stops in OmniLanguage.  As an example, the following stop fires an exit signal if the advisor falls below a certain fixed value.

OmniLanguage sample code of a stop that uses advisor values to determine exits.

Similarly, calls to LongAdvisor() and ShortAdvisor() can be used in designing trade plans.

Example of usage of advisor values in an OmniScript Trade Plan condition.

Downloads Available for Members Now.

The Strategies discussed in this document are now available in the Downloads section of the Nirvana Club web site and were presented at BASH 2007.