关于statsmodels包Unobserved Components系数估计及外生变量估计逻辑的技术问询

Great questions about the Unobserved Components (UC) model in statsmodels—let's break them down one by one:

Statsmodels implements the UC model using a state space framework, where all parameters (including those for unobserved components and exogenous variables) are estimated together via Maximum Likelihood Estimation (MLE) paired with the Kalman Filter. Here's a practical breakdown:

• First, the model is structured as a state space system: unobserved components like trend, seasonality, and cycle are treated as "state variables" that evolve over time, while exogenous variables (if included) are built into the observation equation.
• The Kalman Filter works recursively through your time series: it predicts the current state (e.g., today's trend value, seasonal component) using past data, then updates that prediction with the actual observed sales figure. Along the way, it calculates prediction errors and their variances.
• These errors are used to build the model's log-likelihood function, which measures how well the model fits your observed data.
• Finally, statsmodels uses a numerical optimization algorithm (default is BFGS) to maximize this log-likelihood. This process spits out estimates for all model parameters: the ones governing trend smoothness, seasonal cycle length, and any exogenous variable coefficients you're interested in.

They are estimated jointly with the trend, seasonal, and cycle components—this is a core strength of the UC model in statsmodels. Here's why that matters:

• When you add exogenous variables (like your marketing spend), they're integrated directly into the observation equation of the state space model. A simplified version looks like this:

  sales_t = trend_t + seasonal_t + cycle_t + marketing_spend_t * β + ε_t
  

  where β is the coefficient for marketing spend, and ε_t is the idiosyncratic error term.
• All parameters—including those controlling trend/seasonality/cycle behavior and β—are estimated at the same time via MLE and the Kalman Filter. This is worlds apart from a two-step approach where you first strip out trend/seasonality from sales data, then regress the leftover residuals on marketing spend.
• Joint estimation avoids the bias that plagues two-step methods: in a two-step process, errors from the first step (estimating trend/seasonality) carry over to the second regression, leading to unreliable estimates of β. By estimating everything together, the model accounts for all sources of variation in your data at once, producing more efficient and unbiased results.

内容的提问来源于stack exchange，提问作者Technologic