Brief
A Gaussian process (GP) predicts a function together with its uncertainty, using a kernel that says how strongly two points in time should be correlated. Implement Cholesky-based GP regression with a composite kernel: a long-term RBF (smooth trend), RBF x periodic (a yearly cycle whose shape can slowly change), rational quadratic (medium-term irregularities) and noise. Hand-derive the gradients of the log marginal likelihood (how probable the data is under given kernel hyperparameters; maximising it tunes them) with respect to every hyperparameter, and check them by finite differences. Fit by L-BFGS or Adam, and produce calibrated predictive intervals (intervals that contain the truth as often as they claim to) that extrapolate the last 10 years of NOAA monthly data.
VICTORY CONDITIONS
- LML >= sklearn's on the same kernel; RMSE <= sklearn's; interval coverage 90-98%; analytic gradients match finite differences to 1e-6 (Held-out RMSE and 95%-interval coverage on the last 10 years; optimised log marginal likelihood (LML)).
- A public repository with a README that shows your result and how to reproduce it.
Terms in plain words
- Gaussian process
- A probability distribution over functions, defined by a kernel, that gives both a prediction and an error bar at every point.
- Kernel
- A function that says how correlated the outputs at two inputs are; combining kernels encodes trends, cycles and noise.
- Log marginal likelihood
- The log probability of the observed data under the GP with given hyperparameters, averaged over all functions; you maximise it to fit the kernel.
- Calibrated intervals
- Prediction intervals whose stated coverage, such as 95%, matches how often they actually contain the true values.
- L-BFGS
- A memory-light quasi-Newton optimiser that uses recent gradients to approximate curvature.
Why it rates ★★☆☆☆
Mauna Loa is the standard GP demo. Hand-derived gradients plus a calibration check keep it from being a copy.