This post was firstly submitted to the Applied Quantitative Investment Management group on LinkedIn. It illustrates a recipe implemented in Python package nnetsauce for time series forecasting uncertainty quantification (through simulation): sequential split conformal prediction + block bootstrap

Underlying algorithm:

  • Split data into training set, calibration set and test set
  • Obtain point forecast on calibration set
  • Obtain calibrated residuals = point forecast on calibration set - true observation on calibration set
  • Simulate calibrated residuals using block bootstrap
  • Obtain Point forecast on test set
  • Prediction = Calibrated residuals simulations + point forecast on test set

Interested in experimenting more? Here is a web app.

For more details, you can read (under review): https://www.researchgate.net/publication/379643443_Conformalized_predictive_simulations_for_univariate_time_series

xxx

Comments powered by Talkyard.