Skip to content

Latest commit

 

History

History
87 lines (64 loc) · 4.12 KB

README.md

File metadata and controls

87 lines (64 loc) · 4.12 KB
Raw

Stock-Price-Prediction-Time-Series-with-NN

Stock Price Prediction using NN,LSTM & CNN

The project predicts the closing stock price based on the last 7 days' data includes opening stock price, highest stock price of the day, lowest stock price of the day, stock volume and closing stock price.

1.Problem Statement

This project aims handle the sequential data of stock market, predicting the closing stock pride every 7 days calculating previously sequential values into account towards predicting the every 7th record in the sequential dataset.

Time Series Dataset | Closing Stock Price Prediction | Sequential Data as Image Representation | Regression Chart

Models : Fully-Connected Neural Networks (FCNN) | Convolutional Neural Networks (CNN) | Long Short Term Memory (LSTM)

Project attemps to learn:

  • Applying and comparing FCNN,CNN & LSTM to Time series data
  • Visualizing the output feature - closing stock prices
  • Visualizing sequential data to 4D image to feed in the CNN model
  • Visualizing sequential data to 3D image to feed in the LSTM model
  • Parameter tuning all the models to get the best results
  • Regression Chart or Lift Chart
  • Previous N days' best value
  • Applying Bi-directional LSTM model
  • Applying best models got for totally another sequential dataset - google company's stock prediction

2. Dataset

The dataset has following 7 features: Date, Open, High, Low, Close, Adj_Close, Volume

  • Removed date and adj_close columns, not using them for prediction
  • Preprocessed the data, dropped the rows with null values, unnecessary columns
  • Normalized numeric data with zscore normalization
  • Normalized the output features as seperate column while used the same which fed as train_y and y_true

Data Visualization

  • The flow of value of closing stock price which we intend to predict

Neural models

FCNN

LSTM

CNN

Comparison

1.Best value of N

LSTM

CNN

2.Bidirectional LSTM Model

  • Tried with different layers and parameter tuning but end up with more RMSE than the regular LSTM
  • Observed that the less layers has relatively less RMSE than the deep layer Bidirectional LSTM
  • Did not work for this problem

RMSE : LSTM - 3.006859064102173 | Bidirectional LSTM - 4.965034008026123

3.Best Models applying to different dataset (Google)

  • Prior performing best FCNN,CNN & LSTM models applying for another dataset to observe the results

RMSE

FCNN - 4.5739

LSTM - 52.1711

CNN - 18.1740

Conclusion

  • Different dataset may have different time series flow causes applying different models on that particular dataset, not the best models for likewise dataset.
  • Bidirectional LSTM should have worked well with time series although the performance increases by minor error rate, need to observe more combinations with it.
  • As we considering previous N days' data to predict the next closing stock value for CNN & LSTM, same approach should have been applied to FCNN to get the fair result, although in this project FCNN beats the others from the first place.

Mini Project 4

Mansi Patel

March 27, 2019

Prof: H.Chen

CSC 215