Close

Stock Price Predictions

Predict stock price mainly using Pension Fund

Project Overview

Developed an advanced predictive model utilizing LSTM neural network architecture to effectively forecast stock price movements. Employed diverse time-series predictive models and conducted in-depth analysis on the intricate relationship between pension funds and stock prices, illuminating their impact on market dynamics. These invaluable insights provided a solid foundation for strategic decision-making. Enhanced the accuracy of our LSTM model by refining time intervals, resulting in an impressive 6% increase and bolstering our forecasting capabilities.

Received a challenging assignment from a prominent company to investigate the correlation between pension funds and stock prices. Employing a comprehensive approach involving sophisticated modeling techniques, trend analysis, and seasonality analysis, our findings revealed that the relationship between these factors is not statistically significant. This research established that pension funds do not serve as a strong indicator for price prediction.

Although our initial objective was to uncover a specific correlation, our successful outcome was to demonstrate the absence of a significant connection between pension funds and stock prices. Our achievements in delivering the desired result earned us the prestigious 2nd award.

Tools Used

  • Python
  • Time-Series Analysis
      • 1. Trend and Seasonality Analysis : Visualized monthly averages of analysis variables and used box plots to analyze seasonality on a monthly, weekly, and daily basis
      2. Correlation Analysis : Measured correlation coefficients to analyze the relationship between stock prices, pension funds, and investment entities. Accounted for spurious correlation by using differenced data from unit root tests.
      3. Unit root Tests : Conducted unit root tests (ADF,PP) to determine the stationary of time series and ensure reliable correlation measurements
      4. VAR Model : Consider interdependencies between variables and determine lag length using information criteria
      5. Cointegration Tests : Employed Johansen test to detect long-term correlations between variables and verify meaningful relationships
      6. Impulse Response Function : Used to access the dynamic response of variables in the VAR model to shocks and determine the relative importance of each variable
  • Time-Series Forecasting
      • ARIMA, SARIMA

      Applied to the Korea Stock Market indices, considering variables such as price, pension funds, and other investors. Feature selection, grid-search, and cross-validation were used to optimize the models' performance

      Machine Learning model - Random Forest

      Predict the fluctuations of KOSPI indices. The dependent variable used was the price difference between the current day's closing price and the previous day's closing price. The independent variables included individual investors, institutional investors, foreigners, and the presence of pension funds. Implemented separate models for each variable combination.

      Deep Learning model - LSTM, Vanilla RNN

      Predict stock prices for different time periods. Vanilla RNN has a tendency for underfitting in smaller data intervals. To address this, LSTM was employed and compared.