Prophet Model - CryptoView Pro

Overview

The Prophet model is optimized for medium to long-term predictions (1 week to 1 month) of cryptocurrency prices. Built by Meta (Facebook), Prophet excels at capturing trends and seasonality patterns. Best for: Weekly and monthly forecasts, trend analysis, seasonal patterns Source: source/models/prophet_model.py

When to Use Prophet

Long-term Trends

Predictions from 1 week to 1 month ahead

Seasonality Detection

Automatically identifies daily and weekly patterns

Confidence Intervals

Built-in uncertainty quantification

Robust to Gaps

Handles missing data gracefully

Model Parameters

The ProphetCryptoPredictor accepts the following initialization parameters:

Parameter	Default	Range	Description
`changepoint_prior_scale`	0.5	0.001-0.5	Flexibility for trend changes (higher = more flexible)
`seasonality_prior_scale`	10	1-20	Strength of seasonality component
`interval_width`	0.95	0.8-0.99	Width of confidence intervals

The default changepoint_prior_scale=0.5 is higher than Prophet’s default (0.05) to handle cryptocurrency volatility.

Prophet Configuration

The model is pre-configured with settings optimized for crypto:

Prophet(
    changepoint_prior_scale=0.5,      # High flexibility for volatile crypto
    seasonality_prior_scale=10,       # Strong seasonality modeling
    interval_width=0.95,              # 95% confidence intervals
    daily_seasonality=True,           # Detect daily patterns
    weekly_seasonality=True,          # Detect weekly patterns  
    yearly_seasonality=False,         # Disabled (not relevant for crypto)
    seasonality_mode='multiplicative' # Better for % changes
)

Seasonality Modes

Multiplicative (default): Seasonality scales with trend magnitude
Better for crypto where volatility increases with price
Additive: Seasonality is constant regardless of trend
Use for more stable assets

Usage Example

import pandas as pd
from models.prophet_model import ProphetCryptoPredictor

# Initialize with custom parameters
predictor = ProphetCryptoPredictor(
    changepoint_prior_scale=0.3,  # Less flexible
    seasonality_prior_scale=15,   # Stronger seasonality
    interval_width=0.90           # Narrower confidence bands
)

# Load historical data
df = pd.read_csv('eth_hourly.csv', index_col='timestamp', parse_dates=True)

# Train the model
metrics = predictor.train(df)
print(f"MAPE: {metrics['mape']:.2f}%")
print(f"Direction Accuracy: {metrics['direction_accuracy']:.2f}%")

# Predict next 7 days (168 hours)
predictions = predictor.predict_future(periods=168, freq='H')

# Access predictions with confidence intervals
print(predictions[['predicted_price', 'lower_bound', 'upper_bound', 'trend']])

Training Process

Data Requirements

Minimum: 100 data points
Recommended: 1000+ hourly data points (6+ weeks) for robust seasonality
Format: DataFrame with close column and DatetimeIndex
Frequency: Works with hourly (‘H’) or daily (‘D’) data

Data Preparation

Prophet requires data in a specific format:

# Internal transformation (automatic)
# Input: df with 'close' column and datetime index
# Output: Prophet format
prophet_df = pd.DataFrame({
    'ds': df.index,      # Dates/timestamps
    'y': df['close']     # Values to predict
})

The prepare_data() method handles this transformation automatically. You only need standard OHLCV data.

Training Steps

Data Preparation (line 53-70)
Converts to Prophet’s ds/y format
Model Fitting (line 88)
Prophet decomposes into trend + seasonality + holidays
Validation (line 92-103)
Generates in-sample predictions for metrics

Performance Metrics

The train() method returns:

{
    'mae': float,                    # Mean Absolute Error
    'rmse': float,                   # Root Mean Square Error  
    'mape': float,                   # Mean Absolute Percentage Error
    'direction_accuracy': float,     # % correct price direction
    'training_points': int           # Number of data points used
}

Prediction Horizons

Horizon	Typical MAPE	Direction Accuracy	Use Case
1-3 days	5-8%	55-60%	Short-term trend
1 week	8-12%	50-58%	Weekly analysis
2-4 weeks	12-20%	48-55%	Monthly outlook

Prophet works best for horizons of 1 week or longer. For predictions under 72 hours, use the XGBoost model.

Predictions with Uncertainty

Prophet automatically generates confidence intervals:

# Predict next 14 days
predictions = predictor.predict_future(periods=336, freq='H')  # 336 = 14*24

# Result DataFrame columns:
# - timestamp: Prediction date/time
# - predicted_price: Point estimate (yhat)
# - lower_bound: Lower confidence bound (yhat_lower)
# - upper_bound: Upper confidence bound (yhat_upper)
# - trend: Trend component only

# Example usage
for idx, row in predictions.iterrows():
    print(f"{idx}: ${row['predicted_price']:.2f} "
          f"[${row['lower_bound']:.2f} - ${row['upper_bound']:.2f}]")

The confidence intervals widen further into the future, reflecting increasing uncertainty.

Frequency Options

The predict_future() method accepts different frequencies:

# Hourly predictions (default for crypto)
predictions = predictor.predict_future(periods=168, freq='H')

# Daily predictions (for longer horizons)
predictions = predictor.predict_future(periods=30, freq='D')

# Other pandas frequency codes:
# '30min' - 30 minutes
# '4H' - 4 hours  
# 'D' - daily
# 'W' - weekly

Use hourly frequency for crypto to capture intraday patterns. Switch to daily for very long-term predictions (1+ month).

Backtesting

Use the backtest_prophet() utility function:

from models.prophet_model import backtest_prophet

predictor = ProphetCryptoPredictor()

# Test on last 168 hours (1 week)
results = backtest_prophet(
    df, 
    predictor, 
    test_periods=168
)

print("Train Metrics:", results['train_metrics'])
print("Test Metrics:", results['test_metrics'])

# Access actual vs predicted
train_actual = results['train_actual']
test_actual = results['test_actual']
test_predicted = results['test_predicted']

# Full prediction DataFrame with intervals
predictions = results['predictions']

Backtesting Process

Split data: all except last test_periods for training
Train Prophet on training data
Predict test_periods into the future
Compare predictions against actual held-out data
Calculate test metrics (MAE, RMSE, MAPE, direction accuracy)

Trend Analysis

Prophet decomposes predictions into components:

# After training and prediction
predictions = predictor.predict_future(periods=168, freq='H')

# Extract trend (no seasonality)
trend = predictions['trend']

# Identify trend direction
if trend.iloc[-1] > trend.iloc[0]:
    print("Upward trend detected")
else:
    print("Downward trend detected")

# Measure trend strength
trend_change_pct = ((trend.iloc[-1] - trend.iloc[0]) / trend.iloc[0]) * 100
print(f"Trend change: {trend_change_pct:.2f}%")

Configuration Examples

Conservative (Stable, smooth predictions)

predictor = ProphetCryptoPredictor(
    changepoint_prior_scale=0.1,   # Less flexible trend
    seasonality_prior_scale=5,     # Weaker seasonality
    interval_width=0.90            # Narrower confidence
)

Aggressive (Reactive to changes)

predictor = ProphetCryptoPredictor(
    changepoint_prior_scale=0.5,   # High flexibility (default)
    seasonality_prior_scale=15,    # Strong seasonality
    interval_width=0.95            # Wide confidence bands
)

Long-term Forecasting (1+ month)

predictor = ProphetCryptoPredictor(
    changepoint_prior_scale=0.3,   # Moderate flexibility
    seasonality_prior_scale=20,    # Very strong seasonality
    interval_width=0.80            # Tighter bounds for stability
)

# Use daily frequency for speed
predictions = predictor.predict_future(periods=60, freq='D')

Understanding Changepoint Prior Scale

This parameter controls how flexible the trend is:

0.001-0.01: Very rigid, smooth trend (ignores volatility)
0.05: Prophet default (too smooth for crypto)
0.1-0.3: Moderate flexibility (good for stable crypto)
0.4-0.5: High flexibility (default, handles volatility)

Crypto markets are more volatile than traditional stocks, so we use 0.5 instead of Prophet’s default 0.05.

Performance Optimization

Training Speed

Prophet uses MCMC sampling (slower than XGBoost)
Typical training time: 5-30 seconds for 1000 data points
Use daily frequency for datasets >10,000 points
Consider downsampling very long histories

Memory Usage

Lightweight compared to XGBoost
Stores trend components and seasonality
Minimal feature engineering (just ds/y format)

Prediction Speed

Fast once trained (~50-100ms for 168 predictions)
No iterative process (predicts all periods at once)
Much faster than XGBoost for long horizons

Advantages Over XGBoost

Direct multi-step prediction: Predicts entire future in one pass (no error accumulation)
Uncertainty quantification: Built-in confidence intervals
Trend decomposition: Separate trend from seasonality
Missing data handling: Automatically handles gaps
Interpretability: Clear trend and seasonal components

Limitations

Short-term Accuracy: Less accurate than XGBoost for <72 hour predictions
Feature Engineering: Cannot use technical indicators (RSI, MACD, etc.)
Training Time: Slower to train than XGBoost
Regime Changes: May lag during sudden market shifts
Volatility: Assumes relatively stable variance (crypto can be chaotic)

When to Choose Prophet

Use Prophet

Predictions >1 week ahead
Need confidence intervals
Want trend decomposition
Weekly/monthly planning
Seasonal pattern analysis

Use XGBoost

Predictions <72 hours
Need highest accuracy
Have technical indicators
Intraday trading
Short-term momentum

Next Steps

XGBoost Model

Learn about short-term predictions (1-72 hours)

Hybrid Model

Combine XGBoost + Prophet automatically

Model Comparison

Compare all three models side-by-side

API Reference

Detailed API documentation

Get Started

Core Concepts

ML Models

Features

Configuration

​Overview

​When to Use Prophet

Long-term Trends

Seasonality Detection

Confidence Intervals

Robust to Gaps

​Model Parameters

​Prophet Configuration

​Seasonality Modes

​Usage Example

​Training Process

​Data Requirements

​Data Preparation

​Training Steps

​Performance Metrics

​Prediction Horizons

​Predictions with Uncertainty

​Frequency Options

​Backtesting

​Backtesting Process

​Trend Analysis

​Configuration Examples

​Conservative (Stable, smooth predictions)

​Aggressive (Reactive to changes)

​Long-term Forecasting (1+ month)

​Understanding Changepoint Prior Scale

​Performance Optimization

​Advantages Over XGBoost

​Limitations

​When to Choose Prophet

Use Prophet

Use XGBoost

​Next Steps

XGBoost Model

Hybrid Model

Model Comparison

API Reference

Build docs developers (and LLMs) love

Overview

When to Use Prophet

Model Parameters

Prophet Configuration

Seasonality Modes

Usage Example

Training Process

Data Requirements

Data Preparation

Training Steps

Performance Metrics

Prediction Horizons

Predictions with Uncertainty

Frequency Options

Backtesting

Backtesting Process

Trend Analysis

Configuration Examples

Conservative (Stable, smooth predictions)

Aggressive (Reactive to changes)

Long-term Forecasting (1+ month)

Understanding Changepoint Prior Scale

Performance Optimization

Advantages Over XGBoost

Limitations

When to Choose Prophet

Next Steps