Adaptive anchoring model: how static and dynamic presentation of time series influence judgments and predictions
van Schaik, P
This is an open access article under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits use, distribution and reproduction in any medium, provided the original work is properly cited.
When attempting to predict future events, people commonly rely on historical data. One psychological characteristic of judgmental forecasting of time series, established by research, is that when people make forecasts from series they tend to underestimate future values for upward trends and overestimate them for downward ones, so-called ‘trend-damping’ (modeled by anchoring on, and insufficient adjustment from, the average of recent time series values). Events in a time series can be experienced sequentially (dynamic mode) or they can also be retrospectively viewed simultaneously (static mode), not experienced individually in real time. In one experiment, we studied the influence of presentation mode (dynamic and static) on two sorts of judgment: (i) predictions of the next event (forecast), and (ii) estimation of the average value of all the events in the presented series (average estimation). Participants’ responses in dynamic mode were anchored on more recent events than in static mode for all types of judgment but with different consequences; hence dynamic presentation improved prediction accuracy, but not estimation. These results are not anticipated by existing theoretical accounts; we develop and present an agent-based model - the Adaptive Anchoring Model (ADAM) to account for the difference between processing sequences of dynamically and statically presented stimuli (visually presented data). ADAM captures how variation in presentation mode produces variation in responses (and the accuracy of these responses) in both forecasting and judgment tasks. ADAM’s model predictions for the forecasting and judgment tasks fit better with the response data than a linear-regression time series model. Moreover, ADAM outperformed autoregressive-integrated-moving-average (ARIMA) and exponential-smoothing models, while neither of these models accounts for people’s responses on the average estimation task.
Petko Kusev, Paul van Schaik and Nick Chater are supported by Economic and Social Research Council Grant RES-000-22-1768. We also thank the Nuffield Foundation (SGS36177) and The British Academy (SG47881/SG091144) for supporting Petko Kusev in his research. We are grateful to Peter Barr for his helpful comments and programming the experiment.
This is the author accepted manuscript. The final version is available from Wiley via the DOI in this record.
Published online 6 April 2017