predictive neural network
graph based
OR
INSTRUCTIONS:
The VisFin model is applicable for time series and is specialized in weekly predictions
of equity price. Unlike most common predictive models it uses a neural network with
convolutional (2d) layers, although maintaining slim structure and uses proper
data processing to transform time-series into graphs. It should be noted that the model
will lose accuracy with extremely volatile assets.
You can test the model by using data from the database. To do that, first,
select an asset form 'use database asset' click on the predict button. Alternatively,
you can supply the model with your own data by uploading a CSV file. For the best result
the
file should be formatted as follows:
1) First column (optional) date, in order for the backend to properly read it, the
format must be YYYY-MM-DD and dates should
start from the most recent date to the oldest observation (i.e. the most current is at
the top).
2) Second column ( or first if you skip date) data, again structured
from most recent to older data points.
*** The model expects labels, even if there are non,
simply add date/data in the first row!
Model information:
The model is built by using Tensorflow and Keras API. In addition to the graphical
nature of the model, it uses categorical prediction, where categories are built on the basis
of the asset's standard deviation. By nature, this is supposed to be a short-term prediction
and uses about three months of historic data.