Exponential Smoothing — Simple, Holt-Winters

Exponential smoothing methods assign exponentially decreasing weights to past observations. More recent data gets higher weight, making these methods responsive to level changes.

Simple Exponential Smoothing (SES)

For data with no trend and no seasonality.

Equivalent Formulation

Holt's Linear Trend (Double Smoothing)

Extends SES to capture trend using two equations.

Forecast

Holt-Winters (Triple Smoothing)

Extends Holt's method to capture seasonality. Two variants exist.

Additive Seasonality

Multiplicative Seasonality

Parameter Optimization

The smoothing parameters are typically chosen by minimizing a loss function.

Common loss functions:

• MSE: (most common)

• MAE: (more robust to outliers)

• MAPE: (scale-free)

Python Implementation

import numpy as np

import pandas as pd

import matplotlib.pyplot as plt

from statsmodels.tsa.holtwinters import ExponentialSmoothing



np.random.seed(42)



# Simulate monthly data with trend and seasonality

n = 120

t = np.arange(n)

trend = 200 + 1.5 * t

seasonal = 15 * np.tile(np.sin(2*np.pi*np.arange(12)/12), n//12)

noise = np.random.randn(n) * 3

y = trend + seasonal + noise



dates = pd.date_range('2015', periods=n, freq='M')

ts = pd.Series(y, index=dates)



# Simple Exponential Smoothing

ses = ExponentialSmoothing(ts, trend=None, seasonal=None).fit(smoothing_level=0.3)

print(f"SES alpha: {ses.params['smoothing_level']:.3f}")



# Holt's Linear Trend

holt = ExponentialSmoothing(ts, trend='add', seasonal=None).fit()

print(f"Holt alpha: {holt.params['smoothing_level']:.3f}, beta: {holt.params['smoothing_trend']:.3f}")



# Holt-Winters Multiplicative

hw = ExponentialSmoothing(ts, trend='add', seasonal='mul', seasonal_periods=12).fit()

print(f"HW alpha: {hw.params['smoothing_level']:.3f}, beta: {hw.params['smoothing_trend']:.3f}, gamma: {hw.params['smoothing_seasonal']:.3f}")



# Forecast

forecast = hw.forecast(12)

print(f"\n12-month forecast: {forecast.round(1).values}")

Worked Example

Key Takeaways

Related Topics

• See ARIMA Models for more flexible time series models

• See Seasonal Decomposition for understanding components

• See Stationarity for testing stationarity before modeling