Quantifying and Visualizing Forecast Uncertainty with the FIFE

Abstract

Survival analysis can be a useful tool for modeling the attrition of service members, particularly when it comes to forecasting future states of survival for those members. Government sponsors are often interested in predicting these attrition rates at future time points. The Institute for Defense Analyses (IDA) has developed a tool for this purpose: the Finite Interval Forecasting Engine (FIFE). FIFE is a forecasting tool that produces predictions with various modeling frameworks, including deep neural networks and gradient boosted trees. FIFE combines methods from both survival analysis and multivariate time series analysis to predict future states of survival, along with total counts of attrition, for service members at various future points in time. We discuss methods for quantifying uncertainty in these survival forecasts, both for individual probabilities of exit, and aggregated total exits. While FIFE currently uses advanced approaches for maximizing forecasting performance, through the use of Light GBM for gradient boosted trees and Keras for neural networks, there are currently little to no implemented methods for measuring uncertainty in these predictions.

Open PDF

Document Details

Document Type
Technical Report
Publication Date
Dec 01, 2021
Accession Number
AD1228865

Entities

People

  • Alan B. Gelder
  • Evan T. Miyakawa
  • James M. Bishop
  • John W. Dennis

Organizations

  • Institute for Defense Analyses

Tags

Fields of Study

  • Computer science

Readers

  • Computational Fluid Dynamics (CFD)
  • Naval Personnel Management
  • Neural Network Machine Learning.

Technology Areas

  • AI & ML
  • AI & ML - Bayesian Inference
  • AI & ML - Neural Networks