Hints & Tips
NB: Many users have problems using intraday data with the GPF. Excel indeed automatically converts date/time to a numerical format. Please take a look at this Excel spreadsheet sample to go around this small 'annoyance': GPF Date/Time
The G.P.F. is very versatile, and can generate a number of winning patterns for each set of parameters and objectives. We found that using individual patterns (or a small group of) can enhance an neural net input set. However, using a large set of patterns even from coherent objectives and parameters won't make a good input set on their own. They are likely to be too highly correlated with one another unless you mix and match objectives (type of objective, and number of days ahead) and features, particularly StoK, pyramiding, and BestDay . Please send us your feedback on this.
We usually generate up to five long and five short GPF patterns, based on the shape of the pseudo-equity curve and the out-of-sample stability.
Following Tip 1, we have decided to use a set of G.P.F. patterns and use such combination in our decision making, using logical rules or fuzzy logic rules. For instance, you run the G.P.F. with a %-Long objective over 1, 3 and 5 days ahead, all other parameters being identical (except maybe for the number of bars back which may also vary). This makes 3 optimizations which can bring a total of let's say 12 patterns as listed on the Analysis form. Signals for each patterns can also be given a weighting according to their listing. You can then use a scoring system to decide of the best all-round signal. We recommend using fitness per quote and / or fitness per signal to rank patterns, either on out-of sample data (if large enough to be statistically significant), or on average between in-sample and out-of-sample.
Using pyramiding (or signal accumulation) generates many more signals. This is due to the fact that pyramiding allows for new signals while already in position. It is recommended to then use the 'minimum gain' parameter, to filter mediocre signals, and/or use the 'Average Signal' objective instead of 'Profit'. When using a scoring system like described above in Tip 2, using filtered pyramiding is recommended.
Pyramiding can be used to build a simple on/off system, i.e. stay in position until the signal is cancelled (i.e. reset).
Switching the pyramiding off will generally yield better quality long or short signal entries. All ignored signals (and equity) are listed in the "Equity Curve" worksheet. It may be worth reminding that adding back the ignored signals and optimizing with pyramiding will return different patterns, as the two processes are simply unrelated. More generally, changing any parameter will cause the genetic optimization process to potentially follow a very different route (that's the beauty of it!). For instance, searching for patterns with a Genetic StoK filter will return different patterns than a process first initiated without that filter, then filtered with one.
Please note that when signals are not allowed to accumulate, only consecutive signals are cancelled.
A default population is large enough in most cases. Only increase it if you feel the G.P.F. did not 'try hard' enough. On most occasions, the 'Peaks' and 'Plateaus' parameters should be first increased. It is not uncommon to have a genetic optimization of around 100 to 150 generations. We usually set the Fitness Peaks parameter to 35 to 50.
Using the Optimal StoK and/or the 'Best Day' features will significantly enhance your optimizations, but will reduce the number of signals. It is important to test your patterns on Out-of-Sample data, and analyse carefully the Significance and Serial Correlation figures. They do provide very good additional signals in a consensus based system with a majority of non-filtered signals.
Do not use too many quotes: the markets have evolved dramatically over the last few years, so have winning patterns. The shape of the equity curve (ideally upward and near-linear) is ultimately your best indication of the pattern stability.
We recommend 1000 to 2000 quotes max + eventually another 500 to 1000 for Out of Sample.
There are basically two ways to use out-of-sample data:
It must be said that Genetic Algorithms do not compare with Neural Nets where out-of-sample analysis is there crucial to model generalization. Genetic algorithms are optimizers, i.e do not 'learn'. A number of users have chosen not to do OoS analysis at all.
This may sound obvious but do not use the G.P.F. on securities that have been plain bullish over the whole period of your data series. Successful trading is meant to beat a 'buy and hold' strategy. If you are a long term investor (or a 'monthly' trader) and bought Microsoft or Cisco in 1990 (and sold end of 1999 of course), there is no system which can beat a simple 'buy and hold' strategy.
If you wish to optimize for the best 'Average Size' signal, you may want to use the outlier filtering which indeed may distort the average return. Yet, in most cases, trading is all a matter of taking advantage of the 'fat-tailed' return distribution particular to stocks. Like most professional traders, we feel it would be defeating the point.
The G.P.F. will find the best patterns for the provided specifications, unless you use early stopping techniques that are too constraining. But this does not mean that there are winning patterns for all stocks. The best patterns may sometimes be just not good enough to trade with. We believe this information is as a matter of fact already very valuable.
Do not get carried away, take the generated signal code into SuperCharts, TradeStation or your favorite charting program and start trading with it without a sound money management technique. The G.P.F. generates signals, not a fully-fledged trading system. The G.P.F. takes many trading parameters in consideration in its optimization including stop loss, which is essential to a good money management. Yet, cutting losses is one thing, letting profits run is another. We have not implemented profit targets or trailing stops. GPF signals are best used in a consensus based system and/or as neural net inputs, then fed into a trading strategy. It would have been too far reaching and would have influenced too much the processes the GPF signals are supposed to feed into.
Having said that, you can indeed use the G.P.F. with a 1-day objective and use the signal as a on-off signal to be in or out of the market, often using the pyramiding option. This is maybe the only valid utilization of the G.P.F. as a trading system.
You are most welcome to send your own hints and tips at : firstname.lastname@example.org
Page last modified:
May 08, 2008