Selection of future events from a time series in relation to estimations of forecasting uncertainty

TL;DR

This paper introduces a neural network-based method for selecting more predictable future events from heteroskedastic time series, aiding practical decision-making by focusing on events with lower forecast uncertainty.

Contribution

It proposes a novel procedure combining mean and dispersion neural network models for a priori event selection based on forecast uncertainty, applicable to diverse and heteroskedastic time series.

Findings

Effective in simulated time series

Successfully applied to Dow Jones index data

Improves decision-making by identifying predictable events

Abstract

A new general procedure for a priori selection of more predictable events from a time series of observed variable is proposed. The procedure is applicable to time series which contains different types of events that feature significantly different predictability, or, in other words, to heteroskedastic time series. A priori selection of future events in accordance to expected uncertainty of their forecasts may be helpful for making practical decisions. The procedure first implies creation of two neural network based forecasting models, one of which is aimed at prediction of conditional mean and other - conditional dispersion, and then elaboration of the rule for future event selection into groups of more and less predictable events. The method is demonstrated and tested by the example of the computer generated time series, and then applied to the real world time series, Dow Jones Industrial Average index.