How a Rolling Random Forest Strategy Predicts Bitcoin’s Weekly Direction

Core Concepts

Random Forest Classifier

Random forest builds many independent decision trees; each tree is trained on a random subset of the data (bagging) and at each split considers a random subset of features, reducing correlation between trees. Final prediction is obtained by majority voting.

Rolling Prediction Evaluation

Rolling prediction simulates performance in a dynamic market by repeatedly retraining the model on a fixed‑size recent window and forecasting a future horizon. After each iteration the window slides forward one day.

Specific Steps

Define a training window (e.g., past 30 days).

Train the model on data within the window.

Predict the price direction for the next 7 days.

Slide the window forward by one day.

Repeat and aggregate predictions to compute overall metrics.

Python Implementation

Setup and Configuration

import pandas as pd
import numpy as np
import yfinance as yf
from sklearn.ensemble import RandomForestClassifier
from sklearn.preprocessing import StandardScaler
from sklearn.metrics import (accuracy_score, precision_score, recall_score,
                             f1_score, roc_auc_score, confusion_matrix)
import warnings

TICKER = 'BTC-USD'
START_DATE = '2021-01-01'
PREDICTION_HORIZON = 7
TRAINING_WINDOW_DAYS = 30
TOP_FEATURES = [
    'ROC_10', 'STOCHRSI_d', 'ADX_14', 'STOCHRSI_k', 'RSI_14',
    'STOCH_k', 'ATR_14', 'EMA_20', 'STOCH_d', 'MACD',
    'ULTOSC', 'BB_upper', 'SAR', 'Open_Close', 'MACD_hist'
]
# Random forest hyper‑parameters omitted for brevity

Data Loading and Target Definition

Historical price data are downloaded with yfinance. Technical indicators are computed for each row. The target variable is set to 1 if the price after PREDICTION_HORIZON days exceeds the current price, otherwise 0.

Rolling Prediction Loop

# --- Rolling prediction loop ---
all_predictions = []
all_actuals = []
all_probabilities = []

for i in range(start_index, end_index):
    # 1. Extract current training and prediction windows
    X_train_window = X_all_features.iloc[train_start_idx:train_end_idx]
    Y_train_window = Y_all.iloc[train_start_idx:train_end_idx]
    X_predict_point = X_all_features.iloc[[predict_feature_idx]]

    # 2. Scale features
    scaler = StandardScaler()
    X_train_scaled = scaler.fit_transform(X_train_window)
    X_predict_scaled = scaler.transform(X_predict_point)

    # 3. Train and predict
    rf_model = RandomForestClassifier(
        # hyper‑parameters here
    )
    rf_model.fit(X_train_scaled, Y_train_window)

    # 4. Store results
    all_predictions.append(rf_model.predict(X_predict_scaled)[0])
    all_actuals.append(Y_all.iloc[actual_target_idx])
    all_probabilities.append(rf_model.predict_proba(X_predict_scaled)[0, 1])

Comprehensive Evaluation

After the loop, aggregated predictions and actuals are used to compute accuracy, precision, recall, F1, ROC‑AUC and the confusion matrix.

Results

Accuracy ≈ 68 % (baseline ≈ 51 %).

Precision ≈ 69 %.

Recall ≈ 68 %.

ROC‑AUC ≈ 0.75.

Confusion matrix [[122 57] / [61 128]].

These metrics indicate a statistically significant improvement over random guessing, with roughly two‑thirds of 7‑day direction predictions correct and balanced precision/recall.

Limitations

The script evaluates predictive ability only; it does not implement entry/exit signals, stop‑losses, or risk management.

Market non‑stationarity may cause degradation of performance on future data.

Practical deployment requires extensive testing, hyper‑parameter tuning, and feature analysis.