top of page
Brandyn_Ewanek_vast_infinite_beautiful_abstract_digital_geometr_a5c881d3-9326-4d66-93db-d7

TensorFlow Stock Prediction Meta Model Development
Brandyn Ewanek

Summary of current stage of development of the stock prediction meta model.

Brandyn_Ewanek_vast_infinite_beautiful_abstract_digital_geometr_deb34aed-1baf-433b-8173-75
Neural Network Stock Prediction - 3 Phase Development

2 of 3 Part Project factoring Economic, historical stock price, and Financial News in an effort to predict the stock market.  The long-term goal is to combine the 3 parts is one meta-model. The following is a short summary of the result up the the current stage of development.

shapelytensorflow.jpg

In studying the use of 400 of economic features we were able to significantly reduced the features down to 20 features.  This model branch finished with an MAE that was quite high of 0.97 but the scale of the target being 0 to 1 means that we highly unpredictive and increased number of features could be used in the final meta model.

shapelytensorflow02.jpg

In this massive functional model in TensorFlow we take in 12 inputs in the form of stock indicators of historical price and predict 4 outputs, with 1 Main Target. On our main target our MAE loss was 0.18.  Alone this branch shows some predicability and implies that we have captured valuable information but is unlikely to be predictive by itself.

Summary of Stock Prediction Meta-Model Develop 

The economic branch shows very poor ability to capture the stock price movement with a MAE of 0.97.  This could be because the number of features were reduced too much and we will consider adding some features back into the final Meta-Model.  

The times series branch of our models shows more ability to capture price movements of the stock with a MAE of 0.18.

We should not judge the inclusion of these branches into our final Meta-Model based only on MAE.  The reason is that economic data is released weekly or monthly and there is would have a hard time capturing the daily movements of stock price.  The economic data sets the stage or environment in which the stock resides and the historical prices give daily price movements inside this environment. The hope is that combining these two results will give our Meta-Model the opportunity to reduce the MAE further but it isn't until the NLP branch for intake of news is complete that our model would have a realistic chance of predicting the stock market.

bottom of page