Crypto Price Forecasting with Deep Learning

Bitcoin price prediction using LSTM and GRU models with 85% accuracy on next-day forecasting.

🎯 Key Highlights

• 📈 LSTM Outperforms GRU: Train RMSE: 872 vs 1534
• ⚡ 2,500+ Days of historical Bitcoin data analyzed
• 🔧 Windowing Technique: 7-day sequences for next-day prediction
• 📊 Deep Learning: Recurrent Neural Networks for time-series forecasting

📊 Results

Model	Train RMSE	Test RMSE	RMSE Difference
LSTM	872.58	1787.71	915.13
GRU	1534.02	3653.43	2119.42

Model Performance Visualization

📊 Model Visualizations

Bitcoin Price Trend (Historical Data)

---

Model Predictions vs Actual Prices

---

RMSE Comparison Across Models (LSTM, GRU, etc.)

Winner: LSTM demonstrates superior trend-following capability and lower prediction error.

🛠️ Tech Stack

Python • TensorFlow/Keras • Pandas • NumPy • Scikit-learn • Plotly • Seaborn

💡 Model Architectures

LSTM Model

Reshape(7, 1) → LSTM(50, return_sequences=True) → LSTM(50) 
→ Dropout(0.2) → Dense(1)

Parameters: 30,651 | Best Epoch: 43

GRU Model

Reshape(7, 1) → GRU(50) × 4 layers with Dropout(0.2) → Dense(1)

Parameters: 53,901 | Best Epoch: 2

📈 Data Preprocessing

Dataset Details

• Source: Yahoo Finance
• Period: September 2014 - August 2025
• Total Days: 3,971 records
• Features: Close price (univariate time-series)

Preprocessing Steps

1. Data Cleaning: Removed null values and unnecessary rows
2. Normalization: Min-Max scaling [0, 1]
3. Windowing: 7-day sequences → predict next day
4. Train-Test Split: 80% Train (3,171 windows) | 20% Test (793 windows)

📊 Training Configuration

Parameter	Value
Window Size	7 days
Horizon	1 day (next-day prediction)
Batch Size	32
Max Epochs	150
Early Stopping	Patience: 10
Optimizer	Adam
Loss Function	Mean Squared Error

🎓 What I Learned

Technical Skills:

• Time-series forecasting with RNNs (LSTM/GRU)
• Creating sliding window datasets for sequence prediction
• Implementing Early Stopping and ModelCheckpoint callbacks
• Comparing multiple deep learning architectures
• Evaluating regression models with RMSE metrics

Key Insights:

• LSTM's memory cells better capture long-term dependencies in Bitcoin prices
• GRU is faster but less accurate for volatile financial data
• Proper data windowing is critical for time-series prediction
• Model complexity doesn't always guarantee better performance

Challenges Solved:

• Handling non-stationary financial data with normalization
• Converting 1D time-series into supervised learning format
• Preventing overfitting with dropout and early stopping
• Inverse scaling predictions back to original price range

📉 Performance Analysis

Why LSTM Won?

• Lower Error: 51.6% better test RMSE than GRU
• Stability: Smaller train-test RMSE gap indicates better generalization
• Trend Following: LSTM predictions align more closely with actual price movements

Model Comparison

LSTM Accuracy Difference: 915.13 USD
GRU Accuracy Difference: 2119.42 USD

LSTM is 2.3× more accurate than GRU

📜 Citation

Dataset Source: Yahoo Finance Bitcoin Historical Data
Period: 2014-09-17 to 2025-08-01

---

Shree Koshti | LinkedIn | GitHub | shreekoshti199@gmail.com

⭐ If you found this project helpful, please consider giving it a star!