Probabilistic forecasting and prediction intervals

Prediction intervals serve the purpose of expressing the uncertainty in the predictions. They are mathematically a bit different from confidence intervals (you can learn about the distinction more here) but they serve the same purpose of enhancing your predictions with probabilistic forecasting to deal with uncertainty.

To obtain 95 and 5 percent prediction interval we compute quantiles from residuals between target and prediction we got for the data used for model training. That way we can ensure the interval is thin for situations where the uncertainty is small (e.g. predicting solar production in the middle of the night) and vice versa (e.g. prediction very far ahead).

We refer to this 95 and 5 percent pair as a 90 percent symmetric prediction interval and it could be translated as "an interval, where we expect that an actual value will sit 90 percent of the time". You can manipulate this confidence level to best suit your business needs depending on the trade-off between the cost of over-forecasting or under-forecasting in your business.

Because the TIM forecast is an estimate of a mean of the random distribution (a so called mean forecast), it might sometimes happen, that it does not lie inside this interval. This is caused by the fact that for some distributions the median lies far from the mean.

800px-Visualisation_mode_median_mean.svg.png

This functionality is available only for RTInstantML calls. Prediction intervals are returned only if requested in the configuration. The prediction intervals are then returned as one of the outputs from TIM. Note that prediction intervals are not returned for points that were predicted with safety models.

confidenceExample.png