From Harris et al. 2018.
- Compare multiple models
- Use time-series data when possible
- Account for uncertainty
- Use sensible predictors
- Model unknowns as random effects
- Forecast and assess for different horizons
- Account for observation uncertainty
- Validate via hindcasts
- Publicly archive forecasts
They also illustrate:
- Simple baseline models such as the long term average or random walk (aka "naive"), which in their data typically outcompete more complex models
- ARIMA time-series models
- A few types of species distribution models (SDMs), including two machine learning approaches (boosted regression trees, random forests)
- Incorporating climate forecasts using CMIP scenarios
- Full pipeline from data to evaluating and reporting forecasts
Their code is archived here:
https://github.com/weecology/bbs-forecasting