Bitcoin-signal-prediction-with-Random-Forest

Project Overview

This project applies machine learning techniques to financial time series, with a focus on predicting buy/sell signals for Bitcoin. The core idea is to use a binary classification model (Decision Trees and Random Forest) to identify trading signals derived from the relationship between short-term and long-term exponential weighted moving averages (EWMAs).

Dataset

Source: Bitstamp

Time Interval: Minute-level data

Period: Dec 2011 – Mar 2021 (≈ 9.24 years)

Total Records: 4,857,377 (risk of overfitting due to high granularity) divided by 4 = 1,214,344 records each.

Main Variables: Timestamp, Close

Process Overview

To evaluate the model on realistic, unseen data, the dataset was divided into four sequential parts:

1. Dataset Split

   • Due to file size limitations and to simulate a real forecasting scenario, the full dataset was split into four chronological parts.

2. Training Phase

   • The third dataset was used to train the Random Forest model.
   • This set covers the period immediately before the test set, ensuring that the model learns from past information.

3. Testing / Prediction Phase

   • The trained model was then applied to the fourth dataset, which contains the immediately following data.
   • Predictions were generated on this unseen portion, simulating a real-world scenario where models are trained on past data and then used to predict future market signals.

4. Goal

   • This setup ensures that evaluation is not biased by data leakage and that the model’s performance reflects its ability to generalize to future Bitcoin price movements.

Methodology

Signal Definition

• Golden Cross (Buy = 1): EWM[Short Term Price] > EWM[Long Term Price]

• A bullish technical pattern that occurs when the short-term moving average rises above the long-term moving average, suggesting upward momentum and a potential price increase.

• Death Cross (Sell = 0): EWM[Short Term Price] < EWM[Long Term Price]

• A bearish technical pattern that occurs when the short-term moving average falls below the long-term moving average, indicating downward momentum and a potential price decline.

Short-term EWM: 50 days (72,000 minutes)

Long-term EWM: 200 days (288,000 minutes)

Models Used

• Decision Trees

• Random Forest (ensemble learning)

Workflow

• Data preprocessing and feature engineering

• Label creation (buy/sell signals)

• Train-test split (80% / 20%)

Model training and evaluation

Cross-validation (4-fold)

Libraries

• Pandas → data manipulation

• NumPy → numerical operations

• Matplotlib → visualization

• Scikit-learn → ML models and evaluation

Model Evaluation

Metrics

• Accuracy Score → fraction of correct predictions

• Precision → TP / (TP + FP)

• Recall → TP / (TP + FN)

• F1-score → harmonic mean of precision and recall

• Confusion Matrix → true/false positives and negatives

• AUC-ROC Curve → trade-off between TPR and FPR

Results

• Overall Accuracy: 94%

• Precision for Sell Signals (0): 74%

• Precision for Buy Signals (1): 99%

Conclusions

• The Random Forest model effectively predicts Bitcoin buy/sell signals.

• High accuracy achieved, but sell signals are harder to classify compared to buy signals.

• EWMA-based features provide a robust framework for detecting trend reversals.

Future Improvements

• Test on other cryptocurrencies or financial assets

• Hyperparameter tuning for Random Forest

• Experiment with deep learning (LSTM, GRU) for time series forecasting

From A.Gatto, L.Mosanghini & S.Spagnuolo collaboration.