Compare and comment on the results and methods (pros/cons/utility) employed in problems 1 & 4.
Discussion
- The non-stationary HMM and K-NN perform similarly in correlation skill for both lead times (March \(1^{st}\) and April \(1^{st}\))
- The K-NN method has higher skill in RPSS than the non-stationary HMM with an ~10% improvement for both lead times
- The K-NN method is more effective for data that is not state-dependent (i.e., the forecast for a given year is not significantly dependent on the state of the previous year)
- The non-stationary HMM is superior when considered a large number of covariates as models considering different combinations of covariates can easily be compared used an objective function like AIC
- A non-stationary HMM is less effective for data where the best HMM results in a small number of states and the majority of the data falls under a single state, as the forecast will be less likely to capture a breadth of values