In this issue:
I. Is Detrending Needed for Currencies?
II. ChaosHunter Models – Selecting Outputs
III. ChaosHunter Models – Using Technical Indicators
IV. ChaosHunter Models – Building Models that Last
V. One way to stop this newsletter
***********************************************************
I. Is Detrending Needed for Currencies?
For many years in many places we have warned against using actual price levels in favor of detrended price changes, %changes, normalized indicators, etc. The exceptions have been when the price levels were relatively close over time, with no huge increases in price. We now believe that in many cases currency prices may qualify as an exception, in spite of the sinking dollar in the last few years. This may be especially true if using small bars and only a few months of history. The Euro, for example , has only gone from about 0.9 to around 1.55 since 2001. These ranges, proportionally, are a far cry from say AAPL (a low of about 10 to a high of about 200 in the same period).
The yen is one exception, but if you divide the prices by 100 (as the futures contract does) you are back in a small range around 1.0.
Moreover, with currencies, the price changes and %changes are very small indeed, possibly making it harder to discriminate them in a neural net sometimes.
Lately we have been doing models using the CME currency futures, both with NeuroShell and ChaosHunter without detrending and without adverse effect. Since our Turboprop2 algorithm does have the ability to extrapolate somewhat, the small increments in the currencies are not impossible to handle. Since we are using 5 or 15 minute bars, and hence no more than 2 to 6 months’ history, we feel pretty comfortable with using actual prices of currencies. We do the same with other pairs used as variables in the predictions or trading strategy.
We have been talking about input variables – we still do not believe you should predict prices in the Prediction Wizard if you are getting trading signals.
***********************************************************
II. ChaosHunter Models – Selecting Outputs
NeuroShell Trader owners who buy ChaosHunter to build trading models start with an advantage. The concept of inputs and output matches the terminology of the Prediction Wizard in the Trader, but there’s a twist in ChaosHunter that needs to be understood in order to build profitable models.
Selecting the right inputs works the same in both programs. Look for a handful of indicators or other instrument data that you believe drive the market for the issue you are modeling. In our experience, raw price data is rarely a good input.
The twist is that each program can use the output value differently. If you choose the open as the output in NeuroShell Trader, you’re asking the neural network to find a relationship between the entire set of inputs and the next bar’s open (or percent change in open as we recommend). Once you have the network’s prediction of market direction, you can make your trading decision.
In ChaosHunter, selecting the open as the output serves a different purpose depending upon the optimization goal that you select.
If you choose one of the statistical optimization goal functions such as maximize R-squared, minimize mean-squared error, maximize correlation or maximize % same sign, ChaosHunter will develop a model that will compute the next value in a time series such as the next open. The end result is a predicted value just like the one you get in the NeuroShell Trader’s prediction wizard, but each program uses a different technique to arrive at an answer. ChaosHunter also gives you a formula for calculating the predicted value.
However, in ChaosHunter if you select one of the trading strategy optimization goals such as buy/sell cutoff or buy/sell true/false, choosing the open as an output does not mean that you’re trying to predict the open. It means that ChaosHunter will use the open as the fill price while building a formula that computes a threshold value for making trading decisions or buy/sell true false signals.
***********************************************************
III. ChaosHunter Models – Using Technical Indicators
ChaosHunter can make use of technical indicators if they are included in the original data file. This means, for example, that you can build a chart in NeuroShell Trader Pro and include any of the Trader’s 800 indicators on the chart. Simply export the data as a comma or tab separated text file and ChaosHunter will evaluate those indicators as possible inputs to the model. However, if the indicators come into ChaosHunter as part of a text file, ChaosHunter accepts them at face value, and cannot optimize the indicator parameters, such as the number of periods in a moving average.
However, if you select any of the technical Indicators included in ChaosHunter, the program will optimize the parameters. To find the internal indicators, select the In/Out icon on the toolbar, followed by the Formula Tab, then look for the Technical Indicators.
You can also apply these indicators to more than one data stream from the original text file. On the left side of the screen, click on the desired technical indicators and on the right side of the screen click on the data streams displayed below “Select potential technical indicator time series.” Note that the max lookback period at the top of the screen only applies if you are using the technical indicators within ChaosHunter.
For more information on using ChaosHunter for trading, check out the ChaosHunter help file. For example, the topic Formula Parameters and Chaos Input gives some pertinent advice that can immediately result in better models.
***********************************************************
IV. ChaosHunter Models – Building Models that Last
The biggest challenge that any trader or investor faces is building models that not only work well in hindsight, but into the future as well. This is a challenge to any kind of model, including investing on fundamentals, because after all, even fundamental investing is about making decisions about historical patterns we have observed. Even Warren Buffett has the problem, because stocks chosen using old values may not act in today’s environment as they would have in the past.
In our NeuroShell product one solution was often to reduce the number of “free variables” in the model, so that the model does not actually learn the “noise” instead of the underlying patterns. Neural nets typically try to use in some way all the info they are given. ChaosHunter does not suffer anywhere near as much from the problem of free variables, because it naturally tends to keep only a few variables, even if you have presented it with too many inputs. So in CH that is less of a worry, but in general, we would give CH only what we feel are the most likely leading indicators so it doesn’t waste time finding the wheat in the chaff. CH also does not suffer much from the problem of too much optimization, in our opinion, given that you have enough data as described below.
The major factor in making models that last with NeuroShell, ChaosHunter or anything else right down to discretionary trading is to develop the model or strategy based upon a large, diverse set of relevant patterns from the past. It is also important to allow plenty of future (“out-of-sample”) time to evaluate how well the model held up. After all, Warren Buffett would not assume that the pattern he is using to pick stocks is a failure after say only a few weeks. So let’s examine what “a large, diverse set of relevant patterns” means.
1. A “large set” is one with say at least 1200 bars, and more is better. If you build a model with less data there are many ways to make profitable models from it, not all of which are robust enough to continue working. As an extreme example consider how foolish it would be if you made all future trading decisions based only upon what the market did today.
2. A “diverse set” is one that has many various types of patterns, so that the formula CH produces will have considered many situations that could occur in the future. We would then call the formula “robust”. Now a large set will enhance the probability of getting diverse patterns captured, but will not guarantee it. So you should manually examine the price curves in your set to make sure it shows lots of rising, falling, and volatile patterns.
3. A “relevant set” is one that contains the patterns that are most likely to occur in the future, as opposed to ancient patterns no longer likely to recur. In the mid nineties you could have purchased almost any Internet stock and made a fortune, but very few would advise that pattern of buying today. Of course, deciding what types of patterns in prices are relevant is very difficult. It might be easier to describe what patterns are NOT likely to be relevant. That would include steep bull market price gains, and post bubble burst bear patterns. As this tip is written on September 4, 2008, we appear to be in a somewhat volatile market not likely to take off in either direction without some powerful news event. So if we were building models with daily bars, we would do one of two things:
a. Start our data in early 2003 after the strong bull and bear markets, or
b. Start our data around 1997 so we pick up not only the bear market, but the bull market as well.
If you are building intraday models, make a similar analysis without going back so far in time.
At the end of the day we cannot give you an exact cookbook of how to choose your data periods, because like all of trading, it is more an art than a science. Experiment and come to your own conclusions about what is best for the stocks or other issues you are dealing with. Keep in mind that not every issue is always predictable (has repeating patterns). Sometimes stocks and markets change based on news, and totally new patterns will appear.
***********************************************************
V. One way to stop this newsletter
It is really easy. Just change your email address and don’t tell us.
***********************************************************