8.6 KiB
8.6 KiB
Practical-4 (Recurrent Neural Network - Google Stock Price Dataset)
Problem Statement: Recurrent neural network (RNN): Use the Google stock prices dataset and design a time series analysis and prediction system using RNN.
Note
Dataset available in Datasets directory. In the code, dataset is downloaded directly from Keras/TensorFlow in 2nd step (Load Dataset)
Pre-requisities
- Install packages using
pip:pip install tensorflow keras numpy pandas matplotlib scikit-learn yfinance(tensorflowrequires Python 3.9 - 3.12)
Steps
- Import Libraries
- Load Dataset
- Exploratory Data Analysis (EDA)
- Visualize Closing Price Over Time
- Preprocess Data - Normalize Closing Price
- Create Sequences for RNN Input
- Build the RNN Model
- Train the Model
- Plot Training vs Validation Loss
- Make Predictions and Inverse Scale
- Evaluate the Model
- Plot Actual vs Predicted Stock Price
- Forecast Next 30 Days
Code
1. Import Libraries:
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import yfinance as yf
from sklearn.preprocessing import MinMaxScaler
from sklearn.metrics import mean_squared_error, mean_absolute_error
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Input, Dense, SimpleRNN, Dropout
from tensorflow.keras.callbacks import EarlyStopping
2. Load Dataset:
# Downloads GOOGL stock data from Yahoo Finance for the given date range
ticker = "GOOGL"
df = yf.download(ticker, start="2018-01-01", end="2024-01-01")
# --- Offline alternative (comment out the yf.download above and use this instead if using local dataset) ---
# df = pd.read_csv('GOOGL.csv', index_col='Date', parse_dates=True)
# df = df.sort_index() # ensure chronological order
# yfinance returns MultiIndex columns — flatten to single level
df.columns = df.columns.get_level_values(0)
print(f"Dataset Shape: {df.shape}")
print(f"Date Range: {df.index.min().date()} to {df.index.max().date()}")
print(df.head())
3. Exploratory Data Analysis (EDA):
print("=== Dataset Info ===")
print(df.info())
print("\n=== Statistical Summary ===")
print(df.describe())
print("\n=== Missing Values ===")
print(df.isnull().sum())
4. Visualize Closing Price Over Time:
plt.figure(figsize=(16, 6))
plt.plot(df.index, df['Close'], color='steelblue', linewidth=1.5, label='Close Price')
plt.title('Google (GOOGL) Stock Closing Price (2018–2024)')
plt.xlabel('Date')
plt.ylabel('Price (USD)')
plt.legend()
plt.grid(alpha=0.3)
plt.tight_layout()
plt.show()
5. Preprocess Data - Normalize Closing Price:
data = df[['Close']].values # use only Close price for prediction
scaler = MinMaxScaler(feature_range=(0, 1))
data_scaled = scaler.fit_transform(data) # scale values to [0, 1]
print(f"Original data range: [{data.min():.2f}, {data.max():.2f}]")
print(f"Scaled data range: [{data_scaled.min():.4f}, {data_scaled.max():.4f}]")
print(f"Total data points: {len(data_scaled)}")
6. Create Sequences for RNN Input:
def create_sequences(data, time_steps=60):
X, y = [], []
for i in range(time_steps, len(data)):
X.append(data[i - time_steps:i, 0]) # window of past `time_steps` days
y.append(data[i, 0]) # next day's price
return np.array(X), np.array(y)
TIME_STEPS = 60 # use past 60 days to predict the next day
# 80/20 train-test split (manual, to preserve time order)
train_size = int(len(data_scaled) * 0.80)
train_data = data_scaled[:train_size]
test_data = data_scaled[train_size - TIME_STEPS:] # overlap ensures test sequences start correctly
X_train, y_train = create_sequences(train_data, TIME_STEPS)
X_test, y_test = create_sequences(test_data, TIME_STEPS)
# Reshape to [samples, time_steps, features] — required format for RNN layers
X_train = X_train.reshape((X_train.shape[0], X_train.shape[1], 1))
X_test = X_test.reshape((X_test.shape[0], X_test.shape[1], 1))
print(f"Training samples: {X_train.shape}")
print(f"Testing samples: {X_test.shape}")
7. Build the RNN Model:
model = Sequential()
model.add(Input(shape=(TIME_STEPS, 1))) # input: sequence of 60 days
model.add(SimpleRNN(units=64, return_sequences=True)) # first RNN layer, passes output to next
model.add(Dropout(0.2)) # drop 20% neurons to reduce overfitting
model.add(SimpleRNN(units=64, return_sequences=False)) # second RNN layer, outputs single vector
model.add(Dropout(0.2))
model.add(Dense(units=32, activation='relu')) # fully connected layer
model.add(Dense(units=1)) # output: single predicted price
model.compile(optimizer='adam', loss='mean_squared_error', metrics=['mae'])
model.summary()
8. Train the Model:
# EarlyStopping stops training if val_loss doesn't improve for 10 consecutive epochs
early_stop = EarlyStopping(monitor='val_loss', patience=10, restore_best_weights=True)
history = model.fit(
X_train, y_train,
epochs=60,
batch_size=32,
validation_split=0.1, # use 10% of training data for validation
callbacks=[early_stop],
verbose=1
)
print(f"\nTraining stopped at epoch: {len(history.history['loss'])}")
9. Plot Training vs Validation Loss:
plt.plot(history.history['loss'], label='Train Loss', color='royalblue')
plt.plot(history.history['val_loss'], label='Val Loss', color='tomato')
plt.title('Model Training Loss Over Epochs')
plt.xlabel('Epoch')
plt.ylabel('MSE Loss')
plt.legend()
plt.grid(alpha=0.3)
plt.tight_layout()
plt.show()
10. Make Predictions and Inverse Scale:
y_pred_scaled = model.predict(X_test)
# Convert scaled predictions back to original USD price range
y_pred = scaler.inverse_transform(y_pred_scaled)
y_actual = scaler.inverse_transform(y_test.reshape(-1, 1))
print(f"Sample predictions (first 5): {y_pred[:5].flatten().round(2)}")
print(f"Actual values (first 5): {y_actual[:5].flatten().round(2)}")
11. Evaluate the Model:
mse = mean_squared_error(y_actual, y_pred)
rmse = np.sqrt(mse)
mae = mean_absolute_error(y_actual, y_pred)
mape = np.mean(np.abs((y_actual - y_pred) / y_actual)) * 100 # mean absolute percentage error
print("=" * 40)
print(" MODEL EVALUATION METRICS")
print("=" * 40)
print(f" MSE : {mse:.4f}")
print(f" RMSE : {rmse:.4f}")
print(f" MAE : {mae:.4f}")
print(f" MAPE : {mape:.2f}%")
print("=" * 40)
12. Plot Actual vs Predicted Stock Price:
test_dates = df.index[train_size:] # align dates with test predictions
plt.figure(figsize=(16, 6))
plt.plot(test_dates, y_actual, label='Actual Price', color='steelblue', linewidth=1.5)
plt.plot(test_dates, y_pred, label='Predicted Price', color='tomato', linewidth=1.5, linestyle='--')
plt.title('Google Stock Price: Actual vs Predicted (RNN)')
plt.xlabel('Date')
plt.ylabel('Price (USD)')
plt.legend()
plt.grid(alpha=0.3)
plt.tight_layout()
plt.show()
13. Forecast Next 30 Days:
n_future = 30 # number of future days to predict
# Seed the forecast with the last TIME_STEPS days of known data
future_input = data_scaled[-TIME_STEPS:].reshape(1, TIME_STEPS, 1)
future_predictions = []
for _ in range(n_future):
pred = model.predict(future_input, verbose=0)
future_predictions.append(pred[0, 0])
# Slide the window: drop oldest day, append new prediction
future_input = np.append(future_input[:, 1:, :], pred.reshape(1, 1, 1), axis=1)
# Inverse scale forecasted prices back to USD
future_prices = scaler.inverse_transform(np.array(future_predictions).reshape(-1, 1))
# Generate business day dates starting from the day after last known date
last_date = df.index[-1]
future_dates = pd.date_range(start=last_date + pd.Timedelta(days=1), periods=n_future, freq='B')
plt.figure(figsize=(16, 6))
plt.plot(df.index[-120:], scaler.inverse_transform(data_scaled[-120:]),
label='Historical', color='steelblue', linewidth=1.5)
plt.plot(future_dates, future_prices,
label='30-Day Forecast', color='orange', linewidth=1.5)
plt.axvline(x=last_date, color='gray', linestyle='--', label='Forecast Start')
plt.title('Google Stock — 30-Day Future Price Forecast (RNN)')
plt.xlabel('Date')
plt.ylabel('Price (USD)')
plt.legend()
plt.grid(alpha=0.3)
plt.tight_layout()
plt.show()
print(f"\nForecasted price range: {future_prices.min():.2f} USD - {future_prices.max():.2f} USD")