Artificial intelligence is a plug-in that
let you build prediction model using neural network and SVM (Support vector
machine) systems.
http://en.wikipedia.org/wiki/Neural_network
To create a prediction model, open the
'prediction model' form (Click on 'AI' then 'Prediction'), in the new form,
click on 'Add'.
The prediction model has many settings; those settings will
help you fine-tune your model.
Each prediction model must have three
periods.
The first period or the learning period is used to train the
model.
The second period or the validation period is
used to choose the best model among all the generated models.
The third
period or the testing period is used to test the model with new data. (Data on
which the model was not trained)
Note that the model is trained only with
data from the learning period.
To change the learning, validation and
testing periods, click on the triangles and move them.
There are two options:
Normal: The first bars (depending on the
learning period) will be associated to the learning samples, and then the next
bars will be associated with the validation samples.
Random: The learning
and validation periods will be merged and the learning and validation samples
will be chosen randomly.
This option let you choose the period that
the prediction item will use.
Inputs are time-series that will be used to
train the prediction item.
This is probably the most important step in the
creation of the prediction item.
Add a new input by clicking on
'Add'.
Columns:
- Select: The select checkbox indicate whether to enable or disable
the current input
- Input Type: Select the input type you would
like to use.
- Input: Display information about the input time-series.
·
Settings: Define settings related to the input
you have selected
·
WS: Let you create multiple inputs from this
input.
Example: if you create 'close' as input and set WS to two, the system
will create two time-series. The first one is 'close' and the second one is
'close one bar ago' or 'ref (close, 1)'.
·
Lag: Let you specify the lag to include for the
input.
Example: if you create 'close' as input and set Lag to 3, the system
will create the following input 'close three bar ago' or 'ref(close,
3)'.
·
Preprocessing: Let you select the pre-processing
calculation that will be applied to the input.
·
R: This button is used to remove the selected
input.
Input Types:
- Symbol Field: You can choose between open, high, low, close or
volume times-series.
- Symbol Return: This creates a formula that returns the performance
of one of these time-series: open, high, low, close, volume.
- Formula: This lets you build your own time-series using the formula
editor.
- Extern Symbol: This lets you create a time-series based on an
external symbol data.
- Database Field: This lets you create a time-series from a field
database.
The output is the time-series that will be
predicted.
You can predict a symbol price, a symbol returns or any other
time-series.
The output has the same settings as
inputs.
The 'WS' column doesn't exists in the output settings.
You
can create only one output.
A neural network model has different
settings that can dramatically improve or reduce the performance of the
prediction model.
A neural network model is composed of one
input layer, one output layer and zero, one or many hidden layers.
Click on
the circles to select a layer.
Network settings:
- Annealing Epoch: The epoch at which the learning rate increase or
decrease is performed
- Jitter Epoch: The epoch at which the Jitter is performed. Jitter is
the process of adding small random noise to weights of all synapses at regular
intervals in order to get the network out of a potential local minimum
- Learning Rate Increase Factor: Let the network learn faster in
order to improve the global error, value should lies between 1.00 and
1.03
- Learning Rate Decrease Factor: Let the network learn slower in
order to improve the global error, value should lies between 0.97 and
1.00
Input layer:
The input layer is the first
layer of the neural network model.
The number of inputs determines the number
of nodes for this layer.
Hidden layers:
Hidden layers are layers
that are always between the input and the output layer.
Output layer:
The output layer is the
last layer of the neural network model.
There is only one node in the output
layer.
The value that leaves this layer is the predicted value.
Layer settings:
· Transfer function: Activation or transfer function in a back
propagation network defines the way to obtain output of a neuron given the
collective input from source synapses.
· Learning rate: Learning rate is one of the parameters that govern how
fast a neural network learns and how effective the training is.
- Momentum:
Momentum term associated with each synapse
represents its tendency to retain its previous weight change. This factor
varies from zero to one.
- Max Jitter: The maximum limit to the random noise added to weights
while Jitter, Jitter is the process of adding small random noise to weights of
all synapses at regular intervals in order to get the network out of a
potential local minimum.
- Weight Decay: Each time the neural network is trained, the weights
of synapses are multiplied by this 'Weight Decay'. Weight decay value is
usually between 0.9 and 1.
The filter lets you create a formula that
will be used to reject certain bars from the learning process.
Example: If
you want to reject bars where the volume was insignificant then type something
like this: 'volume < 100'.
Select the symbols that will be used in the
learning process.
Specify when to stop the training, there are
three options:
- By error value: Stop when the error value reaches a certain
level.
- By error change: Stop when the error change goes below a certain
value for a determined number of iterations.
- By iterations: Stop after a specific number of iterations.
Select the model that will be used in
prediction among all the models created during the training.
Type: Select the model based on one of these
values:
Network Error:
- MSE (Root mean square error)
- POCID (Directional accuracy)
- POCD (Prediction sign accuracy)
- UTHEIL (Coefficient U de Theil)
Track on set: Choose whether to select the
best model among the training models or the validation models.
In the 'Prediction' form, select an item
then click on 'Train', the 'Prediction Progress' form appears.
The blue line refers to the iteration that
produced the best prediction model depending on your settings.
In the 'Prediction' form, select an item
then click on the small arrow next to 'Predict'.
You have the choice to
choose between 'Predict' and 'Predict a value'.
This option gives you the ability to predict
values for a range of dates.
This option gives you the ability to predict
a value for a specific date.
You have the possibility to re-train a
prediction model on new data. To do so, click on 'Re-Train' button in the
'Prediction' form.
Re-Training Settings lets you choose the
range of dates and the learning and validation samples that will be used for the
re-training.
In the 'Prediction' form, select an item
then click on 'Update'.
You will lose the data of the prediction
model if you update the latter.
Click on 'Reinitialize Model' in the
'Prediction' form to delete the date of the model.
Two functions let you access a model data
from a formula:
|
Predict("category
name",
"model
name") ; |
Return a time-series that contains the
prediction values.
|
PredictSymbol("category
name",
"model
name") ; |
Return '1' if the prediction model was
trained using the current symbol, otherwise returns
'0'.
