|
Trade FOREX with NeuroShell Trader
by Marge Sherald, CEO
|
|
Use Adaptive Net Indicators in Trading Rules Neural networks are well known for their ability to learn patterns in data, but as markets change the question arises as to when does a model have to be retrained to be effective in current market conditions? That problem may be easily solved by using the Adaptive TurboProp 2 (AT2) add-on for NeuroShell Trader.
Neural nets that update in real time
For example, you want to trade the AUS/USD pair and believe that other related currencies plus the price of gold affect the pair. However, you don’t know the exact rules for the relationship. This is a perfect problem for a neural network to solve because it can find patterns in the data that your brain can’t perceive when dealing with multiple inputs over a long period of time. The advantage of AT2 is that you can set it up to automatically retrain with every new bar.
|
|
|
Beyond Classification: Automated Discovery
by Wm. Bruce Weaver, Ph.D.
Director, Monterey Institute for Research in Astronomy
|
|
Like all sciences, astronomy starts with classification. In the field of stellar astrophysics, this first step is the classification of stars by the appearance of their spectra. Early in the 20th Century, after the advent of quantum mechanics, it became clear that the original alphabetic order was physically incorrect, and the temperature classes were reordered into O, A, B, A, F, G, K, M (Oh Be A Fine Girl, Kiss Me), and a second classification dimension, stellar surface gravity, was also understood to have a strong affect on the appearance of the spectra. It turns out that almost all stars fall into these neat categories, which, in 1943, were put on a robust system based on morphological standards.
But there are a lot of stars in the sky, and if you want to understand the age, structure, and dynamics of our Milky Way Galaxy, and how, why, and where it forms stars, you’d like to classify hundreds of thousands of stars. So, for over 50 years, astronomers sought to find ways to automatically classify stars. None were found that were as good as a human expert peering through a microscope at the photographic spectra.
It occurred to me in 1989 that artificial neural networks (ANNs) could be the long-sought after technique for solving this classification problem. Before embarking on the whole, multidimensional problem, however, I decided to test ANNs with much more restricted problems. I was in for one of the biggest shocks of my professional life!
Click here for more on star discovery.
|
|
|
AME Strategies, One Year In
by Philippe Lonjoux, Noxa Analytics
|
|
A user asked an interesting question; does the Adaptive Mixture of Experts (AME) work on Gold ETFs? I believed he was referring to our Volatility-ETF strategy.
My off-the-cuff response was that it might not, unless the VIX proxy was replaced by the proper index which the Gold ETF tries to mirror.
But I had never given the question much thought and felt it deserved some ink.
The graph above shows the result of going long/shortGDX (Market Vectors Gold Miners ETF) according to the very same signals provided by our Volatility-ETF strategy (VIX as proxy, Window=16).
To my surprise, the strategy has put up impressive numbers: a 68.1% annualized return with 57% days correct and a W/L ratio of 1.36. Results are similar on other Gold ETFs (GLD, IAU, UGL, DUST…).
Clearly, my off-the-cuff response was not confirmed. I am still shocked at how closely correlated Gold products are with the market volatility index.
We will be publishing soon on our website a more complete report on the subject. So stay tuned.
Shameless self-promotion: to see Noxa’s own approach to timing the VIX index, check out ourVolatility-ETF strategy.
You may want to take advantage of our current promotion: get a 25% refund from your purchase of NEI, CSSA and AME. This offer is extended until October 15th, 2013.
Happy trading.
Philippe Lonjoux Noxa Analytics, Inc
|
|
|
ChaosHunter Finds Net Inputs
by Marge Sherald, CEO
|
|
ChaosHunter can create readable formulas to model your data using the inputs and functions that you select. You can save a ChaosHunter model and import it into the NeuroShell Trader so it may be used in combination with other trading rules or neural network predictions.
However, sometimes I don’t want to use the entire formula found by ChaosHunter in my Trader model. Instead I want to use the different optimization technique in ChaosHunter to search for the best inputs to a neural network that I later create in NeuroShell Trader.
I’ll give you can example. I exported the data from the NeuroShell Trader FOREX model described above and opened the file in ChaosHunter. I had added a %change in open 1 period indicator on the Trader chart so I could use that value as an output in ChaosHunter.
Next I loaded the Steve Ward Intraday model template from the ChaosHunter File Menu, Load Template option. The template saved me from individually changing the optimization model settings that he recommended in the ChaosHunter help file. I removed the option to use technical indicators and the chaos variable because I wanted ChaosHunter to find a fairly simple model.
I let the model optimize for a short period of time on 12 cores, taking advantage of ChaosHunter’s ability to use other computers on my network. As optimization progressed I dismissed the formulas that included only one of the three Aussie cross pairs and the price of gold from the Trader model. When ChaosHunter found a model with reasonable results that used the AUD/GBP, AUD/EURO and AUD/CAD cross pairs, I stopped optimization.
I returned to NeuroShell Trader and fed those three cross pairs into an Adaptive TurboProp3 indicator that was included in a second Trading Strategy. The Adaptive TurboProp indicator was set to retrain on every bar.
The result was an improved Trading Strategy2 that increased profits compared to the previous Trading Strategy while still being able to keep up with changing markets.
The ChaosHunter data and model files are available from the ChaosHunter section ofwww.ward.net.
|
|
|
|
Neural nets that update in real time continued I created an example that included the Adaptive TurboProp2 indicators as part of the Trading Strategy. The Long Entry condition is buy when the predicted output of the neural network falls below a threshold value. The Short Entry condition is buy when the Predicted output rises above a threshold. I wanted the Trader’s optimizer to find both thresholds.
This is a contrarian model based on the presumption that the market will return to previous levels. A prudent trader would trade this model with appropriate money management stops in mind, but they are not included in this example.
Inputs to the network included other cross pairs with the Aussie dollar: the Euro, British Pound, and Canadian dollar. The Canadian dollar is included because Canada has a resource based economy similar to Australia. I added gold as the fourth input because it is one of Australia’s major exports.
The hourly model trades 10,000 units and pays 3 pip commissions in and out. Margin is set at 25%. Adaptive TurboProp2 includes parameters for how many previous bars to include in the training data and how far ahead you want to predict. I wanted the optimizer to find these values. I linked these parameters on the long and short side in order to create a model that can learn to trade both long and short and reduce the chances of overfitting trending market data. I set the network to retrain on every new bar, which means that the oldest bar is dropped from the training data set as a new one is added.
The result is a model that found some major turning points and consistently turned a profit. The chart for this example is available from www.ward.net in the Adaptive Turboprop2 Examples section. You must own the Adaptive TurboProp2 Add-on to load the chart.
|
|
