📈 Google Stock Price Prediction Using LSTM

Project Overview

This project applies deep learning to perform short-term stock price prediction using Google's historical stock data from August 2004 to December 2024. Beyond just modeling, this project demonstrates exploratory financial analysis, data-driven decision-making, and LSTM model optimization for time series forecasting.

The goal is to:

• Analyze long-term and short-term price trends
• Understand volume dynamics and investor behavior
• Extract insights from technical indicators (e.g., Moving Averages)
• Build, train, and compare multiple LSTM models for short-term forecasting

---

Dataset Summary

• Source: Downloaded via yfinance for Google stock (GOOGL)
• Time Span: August 19, 2004 to December 17, 2024
• Frequency: Daily (excluding weekends/holidays)

Key columns:

• Close, Adj Close, Volume
• Derived features like SMA (10, 30, 60, 90) and log returns

---

Key Insights from Data Analysis

• Weekday Return Analysis: Minimal differences across weekdays → excluded from modeling.
• Close vs Adj Close:
  • Close used for modeling
  • Adj Close used for visualization due to dividend/split adjustments
• Volume:
  • Declining as price rises → suggests maturing/stabilizing stock
  • Initial IPO years showed high trading activity
• SMA Analysis:
  • SMA(10) tracks short-term movement best
  • SMA(30+) gives long-term smoothing
  • Crossovers helped identify momentum shifts

---

Why MA(10) Was Selected

• Best balance between responsiveness and noise reduction
• Helped identify short-term trends
• Used to choose the LSTM time window (T=10), even though not directly fed as a feature
• Will consider including it explicitly in future model iterations

---

Deep Learning Models: LSTM Family

Used Close price only, scaled via StandardScaler. Data prepared with a T=10 sliding window for supervised learning.

Base Model

• LSTM(16) → Dense(1)
• 50 epochs, batch size = 16
• Designed to avoid overfitting on 5K-sized dataset

Tuned & Stacked Models

• LSTM(32), LSTM(64), LSTM(128)
• LSTM(64) with regularization
• LSTM(128) performed best overall

---

Evaluation Metrics

All models evaluated using:

• RMSE: Root Mean Squared Error
• MAE: Mean Absolute Error
• R²: Coefficient of Determination

📋 Model Performance

Model	RMSE	MAE	R²
Base Model (16)	6.04	4.21	0.62
LSTM (32)	5.14	3.61	0.73
LSTM (64)	5.10	3.64	0.73
LSTM (128)	4.87	3.65	0.75
LSTM (64) - Reg	9.63	7.43	0.04

✅ LSTM(128) yielded the lowest RMSE and highest R².

---

Visual Results

• Predicted vs. Actual stock prices (train/test)
• Evaluation of learning stability with loss plots
• Visual trends aligned well with model quality

---

Tools & Libraries

• yfinance – data collection
• pandas, numpy – preprocessing & transformation
• matplotlib, seaborn , plotly express– visualizations
• scikit-learn – StandardScaler, metrics
• tensorflow.keras – LSTM, callbacks, optimizers

---

Final Thoughts

This project reflects:

• Financially-informed feature engineering
• Well-structured model comparison
• Real-world understanding of stock dynamics
• Practical application of LSTM for forecasting

---

Future Improvements

• Add indicators like RSI, MACD
• Explore GRU, BiLSTM, and Transformers
• Apply dropout and early stopping
• Hyperparameter tuning with KerasTuner
• Include macroeconomic/news sentiment data

---

👩‍💼 Author

Project developed by Shiva Dorri as part of her deep learning journey in financial forecasting.

---

Techniques used: LSTM Model Comparison RMSE R² StandardScaler Sequential Modeling Time Series Forecasting Google Stock Analysis Deep Learning TensorFlow