By the end you will be able to
- 1Write the AFNS model in state-space form: a state equation for the factors under P and a measurement equation for the yields.
- 2Run the Kalman filter's predict and update steps, and read the update as a regression of the factors on each new yield (M-D03: Linear regression).
- 3Explain the likelihood, how missing yields are handled, and why the factors are less certain before 2016.
- 4Read the estimate on 204 month-ends (RMS 5.56bp), check that its P-dynamics are stationary (M-E04: Eigenvalues and positive definiteness), and explain why ten years alone cannot pin them down.
Why it matters
L5.4.04: Nelson-Siegel and Svensson fitted each month on its own and then looked at the factors as time series: two steps, with the first step's errors ignored by the second. L5.4.05: Arbitrage-free Nelson-Siegel wrote down one model for the curve; add how its factors move under P and it has eleven numbers. To estimate them on seventeen years of curves at once, with the long end missing for the first seven, we need a way to track factors we never see from yields we sometimes do. The Kalman filter does exactly that, and it hands us the likelihood as it goes.