RNN - Stock Prediction Model using Attention Multilayer Recurrent Neural Networks with LSTM Cells

Author: Soumyendu Sarkar

Data Preparation

• Data Source research and identification (free part of QuantQuote database )
• Importing decade long stock data of the S&P 500 companies
• Cleaning and Normalizing data with zero mean, unit variance and logarithmic scaling for normal distribution
• Processing and Separating data into Input Data (the intra-day price fluctuations) and Expected Output Data (discrete categorized classification values for price gain over consecutive days )
• Forming Data Frames for Training and Testing for Neural Network

Code for Two Distinct Neural Networks for Comparison

• Code for Fully Connected Feed Forward Neural Network classifier using low level Tensorflow Framework API
• Code for Attention Adaptation of Multilayer (2x) Recurrent Neural Networks (100x NN lookback) with LSTM / GRU Memory Optimized Cells using low level Tensorflow Framework API. This follows latest publication and enhancement in Recurrent Neural Networks.
• Accelerated Linear Algebra optimization for faster code execution with XLA JIT (Just in time compilation) directives
• Both these codes uses low level Tensorflow API to facilitate usage of advanced Tensorflow framework features with embedding and model structure refinements

Model Visualization and Tensorboard Embeddings

• Tensorboard code embedding for Graphical Model Visualization and Data visualization

Analysis and Result Visualization

• Diagnostics and Evaluation with Graphical presentation of the effectiveness of the Model
• Confusion Matrix and Accuracy, Precision, Sensitivity and Specificity
• The Graph demonstrates the effectiveness of both the Neural Networks in making daily trading decisions, scored against random buys and sells
• Several measures of effectiveness in decision making