Three Time Series Forecasting Methods Explained Simply

Time series forecasting is the task of predicting future values based on historical data. Examples span industries like weather forecasting, sales projections, and stock price predictions. Recently, it has been applied to forecast trends in cryptocurrencies like Bitcoin. Given its widespread applications, every data scientist should understand the available methods for time series forecasting.

Core Time Series Forecasting Methods

1. Autoregressive Moving Average (ARMA)

• How It Works: Combines past values (autoregression) and white noise (moving average) to predict future values.

• Limitations:

  • Assumes stationary data (constant statistical properties over time).
  • Fails to capture seasonality.

2. Autoregressive Integrated Moving Average (ARIMA)

• Improvement Over ARMA: Introduces differencing to handle non-stationary data.

• Parameters:

  • p: Lag order (past values).
  • d: Degree of differencing.
  • q: Moving average window.
• Limitations: Still struggles with strong seasonal patterns.

3. Seasonal ARIMA (SARIMA)

• Key Feature: Extends ARIMA to model seasonality with additional parameters (P, D, Q, m).
• Use Case: Ideal for data with recurring patterns (e.g., monthly sales peaks).

Practical Example: Predicting Bitcoin (BTC) Prices

Step 1: Data Preparation

• Import BTC historical data using Pandas.
• Focus on closing prices for modeling.

• Split data:

  • Training: Pre-November 2020.
  • Testing: Post-November 2020.

Step 2: Model Implementation

ARMA Model

from statsmodels.tsa.arima.model import ARIMA  
model = ARIMA(train_data, order=(1, 0, 1))  
results = model.fit()  
predictions = results.forecast(steps=len(test_data))  

• Result: High RMSE (inaccurate trend capture).

ARIMA Model

model = ARIMA(train_data, order=(2, 2, 2))  

• Improvement: Lower RMSE vs. ARMA by adjusting differencing (d=2).

SARIMA Model

from statsmodels.tsa.statespace.sarimax import SARIMAX  
model = SARIMAX(train_data, order=(2, 2, 2), seasonal_order=(1, 1, 1, 12))  

• Challenge: Requires hyperparameter tuning for optimal performance.

FAQs

Q1: Which model is best for seasonal data?

A: SARIMA is explicitly designed for seasonal patterns.

Q2: Why does ARMA fail with non-stationary data?

A: It assumes constant variance and mean—invalid for trends or seasonality.

Q3: How to improve ARIMA/SARIMA accuracy?

A: Use grid search to optimize parameters (p, d, q, P, D, Q).

Conclusion

• ARMA: Simple but limited.
• ARIMA: Handles non-stationary data.
• SARIMA: Captures seasonality.

👉 Master advanced forecasting techniques here

Pro Tip: Always validate models with metrics like RMSE and visualize predictions against actuals.