Creating Model

When you click the CREATE MODEL button or menu item, you will open the creating model wizard, which helps you to create a profitable model and prepare it for training and forecasting.

Data Source.

You can select the type of the data you wish to use for creating the model, and updating it during operation. There're 3 data sources available:

1. Yahoo server. When using this data source, historical quotes data for selected symbol is downloaded from Yahoo! finance server.

2. Metastock Files. This data source supposes that you have Equis Metastock database stored at your PC, with historical quotes. When you select this data source, ticker name field is changed so that you will be able which data record to use for building your model. 

In the left pane, select the folder which contains Metastock database. When you select the folder, right pane will be filled out automatically, with the list of available symbols and companies' names for these symbols. Select the symbol, and click OK button. Selected DB record will be used for building the model, and further updating of the model when you click UPDATE button.

3. CSV Files. This option will allow you to use Comma-Separated-Values data files for building the models and updating them. CSV file is the most common data file format, with the values separated by commas. Stock NeuroMaster uses fixed CSV file format, so, you may need to rearrange your CSV data to be compatible with Stock NeuroMaster.

The format of the file which can be used for importing data and building the models is:

Date, Open, High, Low, Close, Volume, Adjusted Close

where Date is in format of:

d-mmm-yy (for example, "8-Sep-06")

Open, High, Low, Close values are floating point values, with point as decimal separator.

Volume is integer value without any separator.

Adjusted close is floating point value, but, IT IS NOT REQUIRED for building the model. Adjusted close value is used for defining stock splits and dividends, and, if you do not have such a field in your CSV data file, it will be duplicated by Close value automatically when importing the data from CSV.

You can find the example of CSV file format which can be imported into Stock NeuroMaster in "Stock NeuroMaster 130/CSV_Example" directory.

Ticker name.

the name of the stock you wish to create a model from. There should be standard symbol name, for example, if you wish to create SP500 model, you should type ^SPC. If your name is not correct, the program tells you so, and advises you to visit the given link for symbol lookup.

Model name

the name of the model with which your model will be saved and used. You can type any name you wish, with any additional info. For example, if you create a model for the SP500, with end of day prices and a small neural network structure, you can name it as SP500_EOD_SMALL. This will help you to operate with your portfolio of created models in further operating steps.

Initial date of downloading history

here you can select which historical range you wish to download and use for training. The standard range is a 1 year history, but you can experiment with different periods.

End Of Day/End Of Week modeling

you can select which trading period you wish to use in creating a model. If you select End Of Week modeling, downloading data will be preprocessed and converted to a 1 week time frame.  

Expected market chaotic character.

This option is used for estimating the chaotic character for the selected ticker, from VERY CHAOTIC to VERY INERT. This parameter determines how many days looking back the system uses for giving one prediction. In unstable markets, it’s recommended you choose ‘VERY CHAOTIC’, and the system will build predictions based on only 1 day back data.

This parameter sets up how far in the past the model will look for building the forecasting. Technically, it means how many deltas will be used for modeling/forecasting, where delta is the difference between current quote and the quote n days earlier. For example, if the model is built to use open, high, low, close historical quotes, delta means current open-previous close, or current high-low price 3 days ago, and so on.

So, Chaotic settings will put 1 day backward deltas into the model-[current price-previous price].

Fast normal settings will put 2 days backward deltas, such as [current price-previous price] and [current price-2 days ago price].

Slow normal parameter sets up 3 days backward deltas

Inert-4 days backward deltas

Very inert-5 days backward deltas

When using deltas, the model looks at the data in the past, and, based on that data, builds predictions. So, in other words, this parameter means how many previous days’ data will be used for building current signal (buy or sell recommendation).

The more the parameter-the more the suggestion that the current price is the result of price movements of previous days. So, if you suggest that the quotes are not changed chaotically for specified symbol, and the changing of trend is able to be determined several days ago (doesn’t matter how, but, previous bars precedes trend change), you may set slow normal, inert and very inert parameter. This way, you suggest that the price turn takes several days for specified ticker.
In case if the stock is not inert, and the price turn does not take more than 2 days, you may set the parameter to chaotic/fast normal.

Noiseness and filtration level.

This parameter selects estimated noiseness of the chosen ticker, and therefore, the filtration level for the price series.  

This parameter describes initial data normalization level. There’re several normalization methods used in the program, but for initial raw data, the program uses sigmoid function for normalization and scaling the data. Normalization is very important point in successful neural network learning process.

So this parameter means how the data will be normalized before feeding neural network structure. No filtration means that the data will be simply normalized by sigmoid function. Other parameters means that the stochastic ‘peaks’ in data will be smoothed. The higher filtration level will give you more peaks/holes smoothing, and therefore, remove the significance of accident prices jumping from neural network modeling. If you feel that there’re many price jumps over the historical period for specified ticker, which has no reason and cannot be determined by looking at the past data, you may set this parameter to some filtration level. The more such a peaks/holes (accident unusual price jumps)-the higher filtration level you may wish to choose. By default, the parameter is set to ‘no filtration’, expecting that there’s no accident major drops/raises of the price in selected historical period.

Dominant trading strategy.

This parameter sets up the strategy of further forecasting.

Desired signals frequency.

Here you can select the number of days between generated forecasting signals, and control the approximate position duration (days between opening position and closing it).  

In data processing module, the principle of waves is used. That means, that based on previously prepared data, there’s additional data calculated named ‘waves’. It is similar to moving averages, but, the resulting data is oscillation of values around PI value. This tangensoid waves are used in neural network modeling and learning process, and therefore, in forecasting. The effect which brings such a data transformation/applying, is that the more waves are used in modeling-the rare signals you will get, following the rules of forecasting signal accepting/declining. So, using 1 tangensoid wave will give you more frequent signals because the weight of the signal from waves’ signal generating rule will be high. Otherwise, if there’re many waves are used in modeling, the signals will be spaced out, and will not be too frequent, as there should be strong signal supporting by every wave in the model.
For easier understanding, we name that parameter as ‘desired signals frequency’, as there’s direct relation between numbers of waves and signals frequency.

Neural Network structure parameter.

This parameter describes which neural network structure will be used for modeling.

There’re 4 options:

1. simple-this mode creates the nn structure with 3 internal layers, 15 neurons in each. The structure does not describe first, input layer, and last, output layer. First layer may have different number of neurons, equal to input data sets which are used for learning, and depends on all other parameters you select in creating model dialog wizard. Output layer may have different number of layers too, and, usually, it has buy/sell/none signals at its outputs, from all signals generating rules used in the model.
2. standard-the internal nn structure is 3 layers, with 30 neurons in each.
3. complex-internal nn structure is 3 layers, with 50 neurons in each.
4. manual-there you can select your own NN structure, if you feel that these 3 above is not optimal for your modeling. You can set up any dimension of neural network, with any number of layers, and any number of neurons in each. For example, if you type 10-30-50-100-500-300, the program will create the NN model, with 6 internal layers, and different number of neurons in each. Large structures take much memory, and the time for their learning will increase in geometric progression.

Usually, large structures are used for very complex data, to determine the dependencies between input data sets and output data sets, when the correlation between input and output data verges towards zero. Also, large structures have more ‘captious’ memory, and therefore, can be used for learing on long historical periods (3 years, and more).

But, there’s another side of using large structures. In case if the data which is used for learning is not too complex, and there’s no need to recognize millions of input combinations to generate correct output, NN structure will not be a stable. So, even after deep learning, it may still generate incorrect output results because the power of such a model does not correspond to simple input/output data correlation and/or number of learning data sets.

If you use 1 year historical data for learning, you may set this parameter up to standard/complex mode, based on other parameters in modeling you select. For shorter periods, It may be successful to use simple structure. And for longer periods, you may select your own structure of neural network manually.

Reserve data for backtesting.

This parameter allows you to select the part of the historical period which will not be used in training, but will be used for estimating a model’s accuracy and profitability. In the case of using the backtest period, the last days are marked as an ‘out of sample’ period. If you select USE ALL DATA, all available data will be used in training.

In creating a model, your computer must be connected to the Internet, in order to download the required data.