Overview

Project Focus:

• Analyze MKR price volatility and predict its price changes using machine learning models

MKR Overview:

• Native governance token of the MakerDAO ecosystem
• MakerDAO supports DAI, a decentralized stablecoin pegged to the US dollar
• MKR holders influence decision-making within the ecosystem

DAI Importance:

• Stability is crucial for decentralized finance (DeFi) applications
• DAI is created by locking collateral in MakerDAO smart contracts

Value of Predicting MKR Price:

• MKR impacts the health of DAI and MakerDAO's protocol
• Accurate predictions provide insights into market sentiment, governance decisions, and DAI stability
• Beneficial for DeFi participants and investors

Prediction Target (y):

• y represents the percentage change in the closing price of MKR over consecutive time periods
• It is calculated as:

y = (close - close_lag_1) / close_lag_1

• close: Closing price of MKR at the current time step
• close_lag_1: Closing price of MKR at the previous time step
• Continuous variable representing the relative change in MKR price
• Positive values = price increase; negative values = price decrease
• Values expressed as decimals (e.g., 0.05 = 5% increase, -0.03 = 3% decrease)

Why Use Percentage Change:

• Normalizes price movements, reducing sensitivity to absolute price levels
• Captures relative price movements, essential for understanding volatility and predicting trends in a highly volatile asset like MKR

Importance of Analyzing Percentage Change:

• Reveals patterns in MKR's volatility and behavior
• Highlights MKR's impact on DAI stability and the MakerDAO ecosystem
• Provides actionable insights for DeFi participants and investors

Features

Crypto Data Fetcher:

• Retrieves OHLC data for selected cryptocurrencies and stablecoins using the Binance and Kraken API
• Includes additional derived metrics and timezone conversion

Stock Data Fetcher:

• Fetches hourly stock data for predefined tickers using Yahoo Finance
• Enriches data with calculated metrics

Feature Engineering:

• Creates various technical features and custom calculations
• Utilizes the TA-Lib library for advanced technical analysis

MKRUSDT Analysis:

• Focuses on the governance token MKR
• Examines factors influencing its price growth
• Uses machine learning models to predict the target variable (y)

Machine Learning Models:

• Implements models such as Linear Regression (LR), Decision Trees (DT), Random Forest (RF), and XGBoost
• Designed to predict MKR price trends

Flask:

• Included for programmatic interaction with the data
• Optional, suitable for deployment

Docker Support:

• A Dockerfile is provided for easy deployment in containerized environments

Datasets

Crypto Data:
Link to Cryptocurrencies Binance Dataset

Link to Cryptocurrencies Kraken Dataset

Stock Data:
Link to Stocks Dataset

Merged Data with Features:
Link to merged Dataset

Structure

stable_coin/  
├── images/                                     # Contains the images that are generated through EDA     
│   ├── boxplot_mkr.png      
│   ├── correlation_matrix_mkr.png   
│   ├── distribution_price_change.png            
│   ├── timeseries_mkrusdt.png      
│   ├── timeseries_eur.png    
│   ├── price_change_correlation_with_volume.png 
├── README.md                      
├── notebook.ipynb/       
│   ├── get_coins                               # Fetches and processes cryptocurrency data     
│   ├── get_stocks                              # Fetches and processes stock data      
│   ├── feature engineering                     # Adds derived metrics for ML models      
│   ├── model evaluation and tuning             # Compares models and saves the best as a pickle file      
├── train.py                                    # Trains the best model            
├── predict.py                                  # Flask application for making predictions       
├── requirements.txt                            # List of required Python packages       
├── environment.yml                             # Conda environment file     
├── LICENSE      
├── Dockerfile                                  # For containerized deployment    

Data Exploration:

MKRUSDT (Maker):
Number of data points: 1862
Close Price:
Mean: 1566.29
Min: 1063.00
Max: 2411.00
Standard Deviation: 298.14

Price Change:
Mean: 0.50%
Max: 6.62%
Min: -4.03%

Volume:
Mean: 572.94
Min: 16.25
Max: 8915.15

7-Day Moving Average (7d_ma):
Mean: 1564.15
Min: 201.77
Max: 2374.29

7-Day Volatility:
Mean: 17.30%
Max: 751.05%

DAIUSD (DAI Stablecoin):
Number of data points: 613
Maximum price change: 0.47%

Correlation for MKRUSDT

Key observations:

• 7d_ma and 30d_ma: Highly correlated with close and open, indicating their importance for identifying price trends
• atr is moderately correlated with price indicators, emphasizing its role in volatility analysis
• Indicators like adx and rsi have weak correlations with price-related variables but useful for providing additional signals
• volume' and volume_change are moderately correlated with certain price metrics, making them valuable for demand-supply analysis
• growth_future_1h and growth_future_24h have weak correlations with other features, suggesting they may be challenging targets to predict directly
• Feature Combination: Moving averages, volatility, and volume-based features form a strong foundation for predicting MKR price trends

Boxplot for Closing Prices for MKRUSDT

Key observations:

• Median price around 1500
• Outliers visible around 2200, indicating occasional price spikes
• Spread and Support Level: Moderate spread within the core trading range, lower whisker extends to ~1000, suggesting a historical support level
• Interquartile Range (IQR): Middle 50% of price activity is relatively concentrated
• Overall Pattern: Indicates a stable trading range with occasional upside volatility
  

Timeseries for MKRUSDT and DAIUSD

Key observations for MKRUSDT:

• Starting Point: Began around $1200 with initial sideways movement until early November
• Upward Trend: Strong rally from November to early December, peaking at ~$2400 in early December
• December Volatility: Multiple peaks above $2000, significant price fluctuations
• Downward Trend: Gradual decline since mid-December, currently trading around $1400 with bearish momentum
• Overall Range: $1000–2400, with most activity between $1400–2000
• Market Pattern: Suggests a completed pump-and-distribution phase

Key observations for DAIUSD:

• Price Stability: Consistent around $1.00, typical for a stablecoin
• Minimal Volatility: Fluctuations mostly within the $0.999–$1.001 range
• Notable Spikes: Brief spike to $1.005 on January 9th, small spike to $1.002 on January 1st
• Peg Stability: Maintains excellent peg stability around $1.00
• Recent Activity (Jan 9–13): Slightly increased volatility, but remains within acceptable ranges

Distribution of Price Change for MKRUSDT

Key observations:

• Distribution Shape: Appears normal (bell-shaped) and centered around 0, indicating balanced price movements
• Most Frequent Changes: Small fluctuations, typically between -1 and +1
• Outliers: A few extreme positive outliers, reaching up to +6
• Tails: Distribution tails extend from roughly -4 to +6
• Peak Frequency: Most frequent changes occur around 250 occurrences for the smallest price movements

Machine Learning Models

Target Variable Analysis: y

Mean y: 0.0557
Standard Deviation y: 5.3460
Key observation:

• Illustrates the frequency distribution of the target variable y across the Train, Validation, and Test datasets
• Majority of values concentrated near 0
• Extreme outliers present, with some values exceeding 5000
• Distribution is highly skewed, with most values in a small range and a few significantly larger ones
• Extreme outliers may adversely affect the model by increasing error and reducing prediction accuracy
• Skewness suggests the model might face difficulty in accurately predicting y

Key observations:

• Interquartile Range (IQR) is small, suggesting that most data points are closely clustered
• Numerous strong outliers exceed 1000
• This aligns with the histogram: the majority of values are small, with a few extreme values
• These outliers can significantly distort metrics like MSE and RMSE during training and validation
• Next Step: further analysis to decide whether to remove or transform the outliers

Transformation of y

• Addressing the skewness and extreme outliers, a logarithmic transformation was applied to y:

y_train_log = np.log1p(y_train)
y_val_log = np.log1p(y_val)
y_test_log = np.log1p(y_test)

Linear Regression (LR)

• Features ppo, trix, atr show positive accuracy drop, meaning removing these features decreases model accuracy
• Features like sma20, cci, and roc show negative accuracy drop, meaning removing these features could improves model accuracy
• Various features have no influence on the accuracy and could be considered for removal

Key Observations:

• Most predictions are centered around 0, with a sharp peak and minimal spread
• Indicates that the model is predicting a narrow range of values, which could suggest underfitting or that the target variable has a limited variance

Decision Trees (DT)

Key observations:

• Higher values for min_samples_leaf (13-15) yield better results, with MSE around 12.9
• max_depth has minimal impact, with stable MSE across depths, except for min_samples_leaf=1, where a significant deterioration occurs at depth=4 (MSE spikes to ~34)
• Optimal Configuration: max_depth=4, min_samples_leaf=13, achieving an MSE of 12.9
• Interpretation: The model benefits from higher leaf sample restrictions, preventing overfitting

Random Forest (RF)

Key observations:

• Best number of trees (n_estimators): 180, achieving an MSE of ~2900.731
• Best max_depth: 10, balancing between underfitting (depth=5) and overfitting (depth=15)
• Best min_samples_leaf: 1, enabling the model to capture detailed patterns
• Final model performance: MSE: 2870.841, RMSE: 53.580, R² Score: 0.085
• While RMSE is reasonable, the low R² score (8.5%) indicates the model has limited explanatory power

Key observations:

• The residuals (in log scale) are mostly clustered around zero, indicating accurate model predictions in this scale
• A few residuals deviate from zero, suggesting areas where the model struggles with accuracy
• No clear pattern in the residuals, indicating that the model is well-calibrated in the log scale
• Despite the well-calibrated residuals in the log scale, the low R² suggests that the model’s explanatory power remains limited

XGBoost

Key observations:

• Majority of points cluster around the diagonal (pink dashed line), indicating model predictions generally align with actual values
• Strong linear relationship between predicted and actual values, suggesting the model captures underlying patterns well
• Most data points are concentrated around the 0 value on both axes
• Data spans from approximately -2 to 6 on both axes, with sparse points in higher value ranges (4-6)
• Some outliers visible, particularly around (-2, 0) and (6, 0)
• Slight tendency to underpredict at extreme values
• Sparsity at higher values: Indicates less reliable predictions in these ranges
• Generating more training data for extreme value ranges (4-6) could improve model reliability in these areas

Key observations:

• Dominant Features: Technical indicators like trix, roc, ppo, cmo, cci, and bop dominate, suggesting strong predictive power for 1-hour predictions
• Time-Based Features: hour, day, month, and year show very low importance, indicating that price movements are more influenced by technical factors than by time (but consider the short time-frame of 3m)
• Cross-Crypto Correlation: Most cryptocurrency tickers (e.g., BTCUSDT, ETHUSDT) have minimal impact, showing limited cross-crypto correlation in 1-hour predictions
• Traditional Market Indicators: Indicators like ^SPX and ^VIX show low importance, with limited correlation to traditional markets
• Fibonacci Levels: Surprisingly low importance across all timeframes, despite their common use in technical analysis
• Focusing on technical indicators is more valuable than time-based or cross-crypto features, while traditional market indicators and Fibonacci levels have limited predictive power

So I considered model simplification by focusing on top 10-15 features and looked at SHAP

Key observations:

• SHAP Analysis: Provides insight into how individual feature values influence predictions
• Each dot represents a data point, with color indicating feature value (blue = low, red = high)
• For "price_change", high values (red dots) positively push predictions, while low values (blue dots) negatively influence them
• Features like cci and roc exhibit a mix of positive and negative effects, suggesting non-linear relationships with the target variable
• Features with narrow SHAP value distributions, like day or ln_volume, have a limited effect on predictions across the dataset

Retraining Result:

Surprisingly, retraining the model with the most important features resulted in lower scores

The best model is:

XGBoost

• Best Hyperparameters:
  • eta 0.250000
  • max_depth 5.000000
  • min_child_weight 1.000000
  • rmse 1.467122
• • MSE on the Validation Set: 2.1524
• • MAE on the Validation Set: 0.0163
• • R² Score on the Validation Set: 0.9247

Installation

• Clone the repository:

  git clone https://github.com/your-repo.git
  cd your-repo

• Set up the environment Using Conda:

  conda env create -f environment.yml
  conda activate your-environment-name

• Using pip:

  pip install -r requirements.txt

Installation Instructions for TA-Lib

• TA-Lib library is required for this project but is not installed automatically via the environment.yaml file
• You need to install it manually due to potential platform-specific compilation requirements

To install TA-Lib, follow these steps:

• Using Conda (Recommended):

conda install -c conda-forge ta-lib

• Using pip: If you prefer pip, ensure you have the required dependencies installed and run:

pip install TA-Lib

• On macOS with Homebrew: First, install the TA-Lib C library:

brew install ta-lib

• Then install the Python wrapper:

pip install TA-Lib

• On Linux: Install the required development library (e.g., for Ubuntu):

sudo apt-get install libta-lib-dev

• Then install the Python wrapper:

pip install TA-Lib

• On Windows: Download and install the precompiled binaries for your system from the TA-Lib website, then install the Python wrapper:

pip install TA-Lib

Make sure TA-Lib is installed before running the application. If you encounter any issues, refer to the TA-Lib documentation for further assistance

How to Use

• Jupyter Notebook (notebook.ipynb):

  • Fetch cryptocurrency and stock market data:
    • Fetches data for cryptocurrencies and stablecoins defined in the coins list.
    • Processes the data and adds derived metrics (e.g., price change).
    • Saves the final dataset as stable_coins.csv.
  • Fetch hourly stock data for predefined tickers:
    • Adds derived metrics.
    • Formats timestamps.
    • Combines all stock data into a single DataFrame and saves it as a CSV.
    • Logs missing or delisted stocks/cryptos as warnings or errors.
  • Perform feature engineering and derive metrics.
  • Evaluate multiple machine learning models.
  • Save the best models as .pkl files.

• Train the Model:

  • Use train.py to train the best-performing model (default: XGBoost) on the processed data.
  • Save the trained model as a .pkl file.

• Deploy the Model with Flask:

  • Use predict.py to deploy the model and provide predictions via Flask.

Flask

• The repository includes a Flask (predict.py) to interact with the trained XGBoost model. The API allows users to predict the 'close' within the next hour.

• Steps to Use:

  • Start the Flask Server
  • Ensure the conda environment is active and run:

    python predict.py --port=<PORT>

• Replace with the desired port number (e.g., 5001). If no port is specified, the server defaults to port 8000

• Example:

    python predict.py --port=5001

• The server runs at http://0.0.0.0:, for example: http://0.0.0.0:5001

Make Predictions

• Send an HTTP POST request with the input features as JSON to the /predict endpoint. Replace with the port you specified earlier

• Example Input:

curl -X POST http://127.0.0.1:5001/predict \
-H "Content-Type: application/json" \
-d '{"ln_volume": -25.422721545090816, "bop": -0.44444, "ppo": 1.0097517730496455}'

• Example Response:

{
  "predicted_growth_rate": -0.002064734697341919
}

Run with Docker

• To simplify deployment, a Dockerfile is provided. To build and run the Docker container:

• Build the Docker image:

docker pull continuumio/anaconda3
docker build -t mkr-coin-analysis .

• Run the container:

docker run -p 5001:5001 mkr-coin-analysis

License

This project is open-source and licensed under the MIT License.