In mlr-org/mlr3forecasting: Forecasting for 'mlr3'

knitr::opts_chunk$set(echo = TRUE)

Current state

Right now The Auto-arima predictions work like this: Each time learner$predict(row_ids = ids) is called the model assumes that ids are the timestamps that come directly after the last training instance (N). The forecasts are always done assuming the next step is N+1 and the number of steps is determindes by the length of ids.

How it should be

Instead of determining the horizon by the length of the testsamples, there should be an argument for this. For learner$predict(row_ids = 10:15, h=3) the following should be done:

1. check if the timestamp according to row 10 is the consecutive one of the last training instance.

2. make h = 3 predictions at timestep 10:

| predicted Timestep | Mean_prediction | SE_prediction | other predictions
| -------- |:--------:| ---:|---:| | 10 | 1 | 0.1 |5 | | 11 | 2 | 0.2 | 6 | 12 | 3 | 0.3 |7 | 3. update the model for timestep 10 without refitting the parameters.

1. make h = 3 predictions for timestep 11

2. and so on

In total, the call learner$predict(row_ids = 10:15, h=3) should produce 6 matrices with each h=3 rows.

Problems and solutions

1. Add a new argument h to task\$truth() and learner\$predict_internal() that specifies the horizon, i.e task$truth(row_ids = 10:15, h=4) should get 6 matrices with 4 rows each.

2. learner\$predict_internal() needs to update the model after the prediction for each row_id.

3. How do we want to handle the multiple matrices?

   • 3-dimensional array?
   • list of data.tables

mlr-org/mlr3forecasting documentation built on June 29, 2023, 11:57 p.m.