Interactive demo · Probabilistic Machine Learning: Advanced Topics · Chapter 18

Forecasting CO2 with a sum of kernels

Monthly CO2 at Mauna Loa, 1958 to 2001. A GP with a sum of four kernels is conditioned on the data before the cut-off year and forecasts the rest: a smooth rising trend (K1, squared exponential), a seasonal cycle that may slowly change (K2, periodic × squared exponential), medium-term irregularities (K3, rational quadratic) and correlated noise (K4, squared exponential). Switch parts off to see what each one does to the forecast and to the marginal likelihood.

Kernel parts

Training data end (cut-off year)

Trend length scale θ1 (years)

Seasonal decay time θ3 (years)

Noise standard deviation σn (ppm)

Try these

The hyper-parameters start at the values that maximise the marginal likelihood of the data up to 1990 (found offline); the four-part kernel and the model are as in Section 18.8.1, with the period of K2 free and a constant mean equal to the training average. The sliders change only the three parameters shown, all others stay at their fitted values. Data: the monthly averages of the 1958–2001 record bundled with statsmodels, a shorter record than the book's. The band is the 95% interval for the latent CO2 level, without the measurement noise.

In the lecture notes