GoldenGem

GoldenGem is a neural network computer program. The default configuration is the standard one, a three level perceptron, which was shown simultaneously, but independently, to be a 'universal nonlinear function approximator'in two articles in the same journal in 1989. K. Funahashi, On the approximate realization of continuous mappings by neural networks, Neural Networks vol 2, 1989] K. Hornik, Multilayer feed-forward networks are universal approximators, Neural Networks, vol 2, 1989]

Freeware but not open source status

The program was supported by user registration, and was converted to freeware in 2006 and2007, first at download.com and then at the main website. Some software websites continue to post information from an earlier pad file relating to a charge for the use of the program, but all links followed will eventually lead to the main site ordownload.com. The software is not open source because there is concern whether editors would have adequate technical expertise.

Operational modes

While it can access the end-of-the-day daily stock, index, and bond information provided byYahoo, MSN and Google; however a serious user would manage text files of his own data in one of two other formats, either a folder of .csv files or a single .txt file downloaded from other sources such as forexrate.com, or the Bank of England, or particular niche commodity prices.The article on Technical Analysis correctly describes the manner in which neural network software can actas a bridge between technical analysis and the more highly regarded fundamental analysis.

ources of confusion

Many users fail to understand the multi-variable nature of the software. A user must provide a set of share prices, indices, interest rates, or exchange rates which he already expects may be useful to predictthe share price of interest. This does not require that the input data needs to be correlated with the share price of interest or with itself. The ideal user already understands basic facts and knows how to make estimates and predictions on his own before his first encounter with the program.

Training and validation

Training is accomplished by the use of a logarithmic sensitivity adjustment. Validation is bya pair of indicator lights. The first indicator light which becomes yellow if both the correlation coefficient and adjusted correlation coefficientof predicted versus actual change is larger than 0.2 and green if it is larger than 0.5 while the second indicator light goes from red to yellow to green as the training input is removed by theuser's control of the sensitivity adjustment. Secondlyvisual comparison of the graphs is needed to ensure the correlation coefficient is not large due to isolated coincidentalsimilarities. Thirdly the user should run a validation data set on an interval in the past.

The adjusted correlation coefficient

The adjusted correlation coefficient is needed because it is possible to obtain a falsely favourablecorrelation coefficient during backtesting by a strategy of returning to the known mean value of pastdata. A neural network will choose this strategy if it fails to find a relationship. Since the known mean value of past data includes time in the relative future during backtesting theuse of the unadjusted correlation coefficient wouldincorrectly reward such a strategy, which would not necessarily be profitableto continue into the future. The adjusted correlation coefficientis the ordinary correlation coefficient multiplied by the variance of the actual changes and divided by the variance of the predicted changes. This becomes small if the neural networkbecomes unstable or if it converges upon a strategy of continual sudden return to the past mean value.The second indicator light shows the minimum of the adjusted and unadjusted correlation coefficients.

The transition function

The transition function is arctan rather than the sometimes used hyperbolic tangent function.The reason is that the arctan function is suitable for an analog neural network.

Limitations

The configuration of the program is limited to analyzing the values of a set of variables that change overtime, with the aim of predicting the future value of one of those variables based only on the current valueof all the variables. Therefore it would not be useful for analyzing credit risk at a single point in time,for example. Also the displays and user interface assume that the data is daily and that no data is providedon weekends. While this is not an essential feature it would make use of the program too confusing to be usedfor intraday data; for example if used for hourly data the tick marks at the bottom of the screen labelled 'weeks' wouldrefer to five hour intervals, and the slider labelled 'days' being set to 21, in an attempt to excludeweekends, would set an interval of fifteen hours.

Whereas the training takes into account the overall relation among the variables throughoutthe entire backtesting interval, there is no particular weight attached to, for example, yesterday's values,or values from last week. A large coincidental change in the values of the variables on the final day of datawould impact upon the prediction. A user can compensate by including what are known as 'stochastics' as input variables. The justification for not doing so is that the use of 'stochastics' would degradethe performance in cases when a change in today's values are genuinely meaningful, and that variability in performance due to the dependence on data from only one day should not affect theoverall average performance of the prediction.

The algorithm is the most widely used and simplest algorithm. Improved algorithms such as conjugate gradient maypossibly be superior.

External links

* [http://www.goldengem.co.uk www.goldengem.co.uk]

Notes