Fish 550 Spring 2023 - 4 Team 2 - Lab 1

[giscus[bot]]

Fish 550 Spring 2023 - 4  Team 2 - Lab 1

https://atsa-es.github.io/fish550-2023/Lab-1/Final_Write_ups/Lab-1-team-2_final.html

[eeholmes]

Nice job and an interesting biological question to ask re predictability based on number of years in the ocean. Good exploration of the stationarity properties and seeing that it was borderline stationary.

Regarding the increased variance of the forecasts, that occurs when there is a random walk component to the model, so ARIMA(p, 1, q) or d=1. Random walks increase in variance with time.

With Kvichak 2.2, you can see one of the problems with auto.arima() and Arima() in general. It has a strong lag-5 cycle but you can't fit a model with just AR(5) so x[t-5] in it. You have to have AR(1, 2, 3, 4, 5) so 4 extra parameters and AIC will penalize that too much. You would have to fit this model manually (wo Arima()) or use a 5-year Fourier cycle as a covariate. The latter you can do with MARSS() but not really with Arima().