Repository in which I put resources that have been helpful for me to learn/master forecasting
- Dr. Franziska Bell's presentation on forecasting
- Start by exploring your data, determine the trend and seasonality
- Two classes of models:
- Classical/Statistical
- ARIMA
- Exponential smoothing
- Machine Learning
- Quantile Regression Forest
- Support Vector Regression
- Recurrent Neural Networks
- Classical/Statistical
- Deciding on which model depends on the amount of historical data available, the correlation with explanatory variables, interpretability constrains and computational complexity.
- You evaluate using backtesting - training up to a timepoint, and then test what happens next.
(More details here) There are two methods:
- Sliding window approach - you take a fixed window of raining data that you move forward at every pass, and forecast the amount in front of it.
- Expanding window approach - particularly useful with less data - you expand the training data from pass to pass, and you test on the ending data.
- There are several metrics one could use, but comparing your model to a naive forecast (assume today's value will hold until tomorrow) is usually preferred.
- You need to estimate the uncertainty (prediction intervals)
- Special events are frequent and heavily affect core metrics. Often with classical statistical methods, adding exogenous variables is not possible. A more in-depth explanation of this can be found here
- Forecasting: Principles and Practice, the classic, authoritative resource on forecasting.