Harvey-Peters Model

The Harvey-Peters model is a state-space time series forecasting model that decomposes a time series into:

• Level
• Slope (Trend)
• Seasonality

It uses a Kalman Filter-based approach to estimate and update these components over time.

Unlike Exponential Smoothing methods, which treat components as weighted averages, the Harvey-Peters model treats them as stochastic latent states.

This means: - Components evolve dynamically over time - Uncertainty is explicitly modeled - Updates are performed using a Kalman Filter

Model Components

1. Level

Represents the baseline value of the series at time t.

2. Slope (Trend)

Represents the rate of change of the level over time.

3. Seasonality

Captures repeating seasonal patterns in the data.

Methodology

• Uses a State-Space representation
• Estimates hidden states using the Kalman Filter
• Continuously updates:
• Level
• Trend
• Seasonal states

Key Characteristics

• Fully probabilistic framework
• Dynamic adjustment of components
• Handles noise explicitly
• Equivalent in structure to Triple Exponential Smoothing but more flexible

Comparison with Holt-Winters (TES)

Feature	Harvey-Peters Model	TES (Holt-Winters)
Framework	State-space	Exponential smoothing
Components	Stochastic states	Smoothed averages
Estimation	Kalman Filter	Weighted updates (α, β, γ)
Flexibility	High	Moderate
Uncertainty modeling	Explicit	Limited

Advantages

• Statistically rigorous framework
• Handles noise and uncertainty well
• More flexible than classical exponential smoothing
• Suitable for complex time series

Limitations

• More computationally complex
• Requires statistical understanding
• Parameter estimation can be harder
• Less intuitive than Holt-Winters models

Use Cases

• Economic forecasting
• Financial time series modeling
• Demand forecasting with uncertainty
• Advanced state-space modeling applications