Intuitive Time-Series Extrapolation

Abstract

The purpose of this paper is to report on a replication and extension of my previous study concerning individuals' ability to intuitively extrapolate from time-series data (Eggleton [1976a; 1976b]). In this study I argued that the accuracy of individuals' estimates of future observations and their associated credible intervals would be a function of their ability to: (1) correctly assess the nature of the underlying data-generating process, (2) cognitively represent the essential characteristics of this process, and (3) generate bias-free predictions from that cognitive representation. In the light of evidence demonstrating limitations to human information-processing capabilities, I hypothesized that individuals would utilize a simple three-stage pattern search, prototype abstraction, and heuristic prediction process which would precipitate systematic prediction errors. To test this and related hypotheses I conducted an exploratory laboratory experiment using a heterogeneous group of 20 adults. The stimuli were 12 sets of 12 two-digit numbers (ostensibly time series of past monthly costs of production) comprising trend, random, and alternating sequences varying systematically in their means and variances. Subjects viewed each sequence for 15 seconds before estimating the next future observation and its associated p = .50, p = .75, and p = .95 credible intervals. After a second 15-second viewing subjects estimated the mean of the sequence.1 The major (tentative) findings were that: (1) Subjects' future cost estimates for trend and random sequences differed, but those