A random walk models a changing quantity by assuming that each new state stays near the current state, with every change coming from random noise. For example, a latent CO₂ level can evolve as xᵢ₊₁ = xᵢ + ε, where ε follows a zero-mean Gaussian distribution. The expected next state therefore equals the current state, while σ² controls how widely the next state may vary. section 9.6
The variance expresses how quickly large changes become unlikely. A small transition variance makes the model conservative; a larger value lets it adapt more readily to observations. In a Bayesian filter, the posterior combines the random-walk prediction with noisy measurements, producing an estimate between the prior prediction and the observed data. section 9.10 section 9.9
Tracking atmospheric CO₂
To estimate the transition variance from observed CO₂ measurements, calculate successive differences and use their empirical variance: load the data with NumPy, compute np.diff(y), and calculate np.var(diffs). A random-walk Kalman filter can then update its estimate as each yearly observation arrives. section 9.6 section 9.9
Random walks are useful when the state may drift but no specific trend mechanism is supplied. They do not actively model systematic long-term movement. Consequently, multi-step forecasts have a flat predictive mean, while uncertainty grows as transition noise accumulates. section 9.15 section 9.11
Use the transition equation to connect the intuitive idea of carrying the current state forward with zero-mean random change. section 9.6



