This repository will consist of a Long Short-Term Memory implementation of a Recurrent Neural Network used to predict the stock prices of Google Stocks for the next working day based on their past few days opening price trends.
- The project is the implementation of Stock Market Price Predicion using a Long Short-Term Memory type of Recurrent Neural Network with 4 hidden layers of LSTM and each layer is added with a Droupout of 0.2 and tested on various values in the Experimentations.
- The Preprocessing involves the Normailzation of the data using MinMaxScaler
- Training set was built considering the past 60 days data of opening stock value as the learning parameters for the LSTM to predict the next day market opening stock value.
- The Data set used in the project is google's stock market data for 5 years starting from January 2012 till December 2016 consisting of various features of stocks like opening value and closing values, etc. which can be downloaded from TRAIN DATA and TEST DATA.
- Various plots for different experimentations are shown below to understand how the model behaves over the given data showing various parameters dependence.
- The first experiment was run for 10 epochs and learned over past 60 days of data to analyze how the model will perform on the data and what could be the kind of predictions we can expect from the low epoch values.
- Below is the screenshot of the real vs the predicted regression for the test dataset.
- The first experiment was run for 25 epochs and learned over past 60 days of data to analyze how the model will perform on the data and what could be the kind of predictions we can expect from the low epoch values.
- Below is the screenshot of the real vs the predicted regression for the test dataset.
- The first experiment was run for 35 epochs and learned over past 60 days and also increasing the droupouts to be 0.05, 0.10, 0.15 and 0.20 for each layer respectively of data to analyze how the model will perform on the data and what could be the kind of predictions we can expect from the low epoch values.
- Below is the screenshot of the real vs the predicted regression for the test dataset.
- The first experiment was run for 50 epochs and learned over past 60 days of data to analyze how the model will perform on the data and what could be the kind of predictions we can expect from the low epoch values.
- Below is the screenshot of the real vs the predicted regression for the test dataset.
- The first experiment was run for 100 epochs and learned over past 60 days of data to analyze how the model will perform on the data and what could be the kind of predictions we can expect from the low epoch values.
- Below is the screenshot of the real vs the predicted regression for the test dataset.
- The first experiment was run for 100 epochs and learned over past 120 days of data to analyze how the model will perform on the data and what could be the kind of predictions we can expect from the low epoch values.
- This is the final experimentation till now before any hyper parameter tuning, as the model is fitting pretty better over the test data.
- Below is the screenshot of the real vs the predicted regression for the test dataset.
- I have achieved an overall loss of around (~0.15%) which is 0.0015 validation loss over 120 days past data training running for 100 epochs, calculated as Root mean squared error calculation.
- Future Scope: The loss can still be optimized based on the Hyper Parameter Tuning to fit the model much better over the testing data.
