Data Science Final Examination

Information

• Student: Bùi Thế Trung - 21110108 - 21110108@st.vju.ac.vn

Problem Statement

• The stock market is volatile and depends on various external factors.
• In addition to external factors, the influence of past stock prices also significantly affects the future.
• Identifying this influence contributes to predicting future stock prices.

Question

• Predict the stock price of Apple company for the next n days (n > 0) based on historical trading data.

Solution

Idea

• Learn about LSTM here.

• Use a Long Short-Term Memory (LSTM) network to analyze the relationship between the past and the future.

• Reasons for using LSTM:

  • LSTM and RNN, in general, are designed to handle sequential data.
  • Each neuron in the LSTM network has three gates and a cell state.

  

  • cell state: This is where a cell can store information (partly related to information processed earlier, partly related to the input data at each cell).
  • forget gate: Decides what information should be discarded or kept (input to this gate includes input data and information from the previous cell).
  • input gate: Filters important information from the input data and the previous cell's information, updating this information into the cell state.
  • output gate: Decides the input for the next cell by further filtering information from input, cell state, and information received from the previous cell.

Execution

• Data Collection

  • Use the DataReader library from Pandas.
  • Install: pip install pandas-datareader.

• Data Preprocessing

  • Keep only the Adj Close Price column for prediction.
  • Scale the data using sklearn.preprocessing.MinMaxScaler.
  • Transform the data into a format that LSTM can read (data in 3D format).
    • sequence: Total input data.
    • time_step: Number of input data needed for one prediction.
    • feature: Number of data samples for one output (calculated from the time_step input data above).

• Build Model

  • Use LSTM from the keras library (install: pip install Keras).
  • The model consists of one LSTM layer.
    • For predicting one day
      • time_step = 60: Use the past 60 days for input data for one prediction.
      • feature = 1: Output data is the price of the next day.
    • For predicting 30 days
      • time_step = 1500: Use the past 1500 days for input data for one prediction.
      • feature = 30: Output data is the stock price for the next 30 days.

• Prediction

  • For predicting one day: the test set contains 2122 values.
  • For predicting 30 days: the test set contains 30 values (if more, handling overlapping predictions is required).

Demo

• Predicting the stock price for the next 1 day (test set includes 2122 values): MSE = 26.652150430735873
• Predicting the stock price for the next 30 days (test set includes 30 values): MSE = 16.084394047657305