Forecast¶

• static / dynamic forecasting

• short term: 2-week lookahead

• long term

data¶

Data aspects:

• missing data,

• trend,

• seasonality,

• volatility,

• drift and

• rare events.

https://neptune.ai/blog/arima-vs-prophet-vs-lstm

methods¶

why does forecast combination work?

• https://scholar.cu.edu.eg/?q=amiratiya/files/m4-ijf.pdf

• average of diverse or comparable forecasts has a shorter distance to the true values compared to each of the forecast

• exclude forecasts that are considerably worse than the best ones in the pool, unless they are very diverse from the rest

retrospect¶

• check accuracy each week

• target: minimize cost or forecast error?

underfit and overfit¶

https://curiousily.com/posts/hackers-guide-to-fixing-underfitting-and-overfitting-models/

• underfit: high bias

• overfit: high variance

• mse (mean squared error) = variance + bias^2

prediction intervals¶

https://otexts.com/fpp2/prediction-intervals.html