Lecture Notes: Forecasting Microsoft Stock using LSTM Neural Networks

Introduction

• Presenter: Greg Hogg
• Focus: Forecasting Microsoft stock using LSTM (Long Short Term Memory) neural networks.
• Importance: Useful project for resumes, recommended to watch the entire video for clarity.

Data Acquisition

• Source: Yahoo Finance, Microsoft Corporation stock page.
• Steps to Download:
  • Change time period from 1 year to max.
  • Download CSV file.
  • Import CSV into Google Colab using pandas.

Data Preparation

• Initial Setup:
  • Import necessary libraries: pandas, datetime.
  • Load CSV data with pandas.read_csv().
• Data Details:
  • 9082 rows of stock data from 1986 to March 23, 2022.
  • Columns: date, open, high, low, close, adjusted close, volume.
• Data Cleaning:
  • Focus on 'date' and 'close' columns.
  • Convert date strings to datetime objects using a custom function.
  • Set 'date' as the index of the dataframe.

Data Visualization

• Plotting:
  • Use matplotlib to plot stock data from 1986 to 2022.
  • Notable trend: significant stock increase post-2016.

Problem Conversion to Supervised Learning

• Method:
  • Create a function to convert dataframe to windowed dataframe for model input.
  • Generate a target date and values for three previous days for each date.

Preparing Data for TensorFlow

• Conversion:
  • Convert data to NumPy arrays suitable for TensorFlow.
  • Separate data into dates, input matrix (X), and output vector (Y).
• Data Splitting:
  • Split data into training (80%), validation (10%), and test (10%) datasets.

Building the LSTM Model

• Tools: TensorFlow and Keras.
• Model Specification:
  • Sequential model with input layer, LSTM layer (64 units), and dense layers (32 units each).
  • Compile using mean squared error as loss function, Adam optimizer.
• Training:
  • Train model on training data, validate on validation data.
  • Monitor mean absolute error for validation.

Model Evaluation

• Performance:
  • Model performs well on training data.
  • Validation and test predictions show poor extrapolation capabilities.

Improvement Strategies

• Re-training with Different Data:
  • Train model on more recent data (post-2021), improve validation/test performance.
  • Adjust data split accordingly.

Recursive Forecasting

• Methodology:
  • Use last known observations to predict future values recursively.
  • Evaluate long-term prediction capabilities.
• Conclusion:
  • LSTM models struggle with extrapolating beyond the trained range.
  • Stock prediction is inherently difficult, especially for specific future dates.

Final Thoughts and Recommendations

• Limitations: LSTM models may not predict well outside trained data range.
• Stock Prediction Advice: Model suitable for long-term trend analysis rather than daily predictions.
• Conclusion: Understanding stock trends is complex, and predictions should be made cautiously.