📈 ProjectML — Bitcoin Price Prediction using Machine Learning & Mathematical Models

A deep-dive into modeling Bitcoin closing prices using classical regression, machine learning, and time series techniques — with Exponential Fit emerging as the top performer.

---

📘 Overview

This project aims to predict the closing price of Bitcoin (BTC) using a mix of traditional regression methods, statistical techniques, and machine learning models. The performance of each model is evaluated using multiple error metrics and k-Fold cross-validation.

To enhance model reliability and generalization:

• I applied Grid Search, Random Search, and a custom Hybrid Search (combining both approaches) for hyperparameter tuning.
• Models were fine-tuned individually for optimal performance.
• A short time series analysis was conducted to identify trend dynamics and stationarity.
• Technical analysis indicators were used as features to help capture market behavior.
• Elements of fundamental analysis and psychological analysis (market sentiment & investor behavior) were considered to contextualize trends and modeling.

---

📊 Dataset

• Source: Yahoo Finance
• Time Range: Weekly data over several years
• Features Used:
  • Open
  • High
  • Low
  • Volume
  • Technical indicators also used for feature selection
• Target Variable: Close price

---

⚙️ Preprocessing & Techniques

• Cleaned and converted to weekly chart
• Applied normalization (Min-Max scaling)
• Feature selection via correlation analysis
• Applied k-Fold Cross-Validation to assess model stability
• Applied hyperparameter tuning via Grid Search, Random Search, and Hybrid Search
• Incorporated time series decomposition and smoothing

---

🧠 Models & Techniques Used

Technical Indicators Used for Features:

• SMA50 (50-day Simple Moving Average)
• RSI
• MACD
• Williams %R

Models Used:

• Linear Regression
• Logarithmic Fit
• Exponential Fit
• kNN
• SMA50
• Random Forest
• SVR
• LSTM

---

📏 Evaluation Metrics

• RMSE (Root Mean Square Error)
• MSE (Mean Squared Error)
• MAE (Mean Absolute Error)
• MAPE (Mean Absolute Percentage Error)
• R² Score

---

🥇 Results

Model	Best Metric Value	Notes
Exponential Fit	✅ Lowest errors & highest R²	✅ Best overall performer
Linear Regression	Moderate	Baseline model
Logarithmic Fit	Decent	Slightly underperformed Exp Fit
kNN	Varies	Sensitive to hyperparameters
Random Forest	Good	Solid performance, but slower
SVR	Fair	Required normalization
SMA50	Baseline smoothing	Not predictive, used for trend only
LSTM	Experimental	Needs more tuning/data

---

🔍 Key Insights

• Exponential Fit captured BTC price growth trends best, outperforming ML models.
• Classical models can rival or beat black-box models on financial time series with the right transformation (especially for assets like Bitcoin with steady exponential growth).
• Smoothing methods (like SMA) help in understanding trends but are poor predictors alone.
• LSTM had potential but needs much more data and tuning for reliable results.
• Combining domain knowledge from technical, fundamental, and psychological analysis adds practical value to the modeling process and improves interpretation of results.

---

🧑‍💻 Author

Created by Panagiotis Mokos
GitHub: @MwkosP

---