|
|
--- |
|
|
tags: |
|
|
- time-series |
|
|
- forecasting |
|
|
- cryptocurrency |
|
|
- lstm |
|
|
- finance |
|
|
library_name: pytorch |
|
|
license: mit |
|
|
datasets: |
|
|
- coingecko_daily |
|
|
metrics: |
|
|
- rmse |
|
|
- mape |
|
|
--- |
|
|
|
|
|
# CryptoPriceForecaster: Deep LSTM for Short-Term Price Trend Prediction |
|
|
|
|
|
## 📑 Overview |
|
|
|
|
|
**CryptoPriceForecaster** is a deep **Long Short-Term Memory (LSTM)** neural network optimized for predicting the closing price trend (i.e., the price change magnitude) of a major cryptocurrency (e.g., BTC/USD) over a short-term horizon (e.g., next 24 hours). It is trained on historical data including price, volume, and derived technical indicators. |
|
|
|
|
|
## 🤖 Model Architecture |
|
|
|
|
|
This is a custom deep learning model implemented in PyTorch/TensorFlow. |
|
|
|
|
|
* **Type:** Recurrent Neural Network (RNN) - LSTM variant. |
|
|
* **Input Features:** 10 features (e.g., Open, High, Low, Close, Volume, 7-day EMA, RSI, MACD Histogram, etc.). |
|
|
* **Sequence Length:** 30 time steps (i.e., the model looks at the last 30 daily data points). |
|
|
* **Layers:** 3 stacked LSTM layers. |
|
|
* `Hidden Size`: 128 units per layer. |
|
|
* `Dropout`: 0.2 applied between layers. |
|
|
* **Output:** A single dense layer predicting the scaled *normalized* closing price. The predicted value needs to be inverse-transformed to get the actual price. |
|
|
|
|
|
## 🎯 Intended Use |
|
|
|
|
|
This model is designed for: |
|
|
1. **Short-Term Trading Strategy:** Generating daily signals for predicting price movement direction. |
|
|
2. **Research:** Studying the non-linear dependencies between technical indicators and price volatility. |
|
|
3. **Educational Purposes:** Demonstrating the use of RNNs for financial time-series forecasting. |
|
|
|
|
|
## ⚠️ Limitations |
|
|
|
|
|
* **Prediction vs. Reality:** This model predicts price based on historical and technical data; it does *not* account for Black Swan events, regulatory changes, or breaking news. **It is not financial advice.** |
|
|
* **Generalization:** Trained primarily on BTC/USD daily data, its performance on altcoins or intraday data may be significantly lower. |
|
|
* **Stationarity:** Requires pre-processed, potentially stationary, input data (e.g., log returns) for optimal performance. |
|
|
|
|
|
## 💻 Example Code |
|
|
|
|
|
Assuming the input data is a 3D numpy array `X_test` (batch_size, sequence_length, features): |
|
|
|
|
|
```python |
|
|
import torch |
|
|
import numpy as np |
|
|
from CryptoPriceForecaster_Model import LSTMForecaster # Custom model class |
|
|
|
|
|
# Load the model weights and config |
|
|
config = json.load(open("config.json")) |
|
|
model = LSTMForecaster(**config) |
|
|
model.load_state_dict(torch.load("pytorch_model.bin")) |
|
|
model.eval() |
|
|
|
|
|
# Example input: 1 sample, 30 sequence length, 10 features |
|
|
X_test_sample = np.random.rand(1, 30, 10).astype(np.float32) |
|
|
input_tensor = torch.from_numpy(X_test_sample) |
|
|
|
|
|
with torch.no_grad(): |
|
|
scaled_prediction = model(input_tensor) |
|
|
|
|
|
# The output must be inverse-transformed using the original scaler (MinMaxScaler) |
|
|
# For demonstration: |
|
|
print(f"Scaled Predicted Price Change: {scaled_prediction.item():.4f}") |
