Algorithm Aversion in Forecasting: A Behavioral Investigation
How to cite
Klöckner, Martin. 2017. "Algorithm Aversion in Forecasting: A Behavioral Investigation." AEA RCT Registry. AEARCTR-0002298. https://doi.org/10.1257/rct.2298-1.0
Download: BibTeX · RIS — for reference managers without a browser translator. Zotero and most others pick up this page's citation metadata directly.
* Primary investigator
Abstract
The research project is positioned within the field of behavioral decision making and focuses on the topic of algorithm aversion in forecasting decisions. Algorithm aversion describes the phenomenon that many individuals tend to rely on human forecasters and discard algorithms even though the latter are generally performing better. Previous research has shown the reasons for this to be manifold and has also discussed possible managerial interventions. However, it has not been examined so far whether the complexity of the forecast algorithm or the reputation of the algorithm developer/provider influence individual's decision to apply an algorithm. This is of particular relevance today as forecasts are becoming more complex in the wake of big data and machine learning technologies. Since forecasts employing such methods are expected to be more accurate than human forecasts and simple statistical forecasting models, failure to make use of them could be very costly for companies. Therefore, this study examines to what extent individuals trust algorithms along the two dimensions complexity and reputation of forecasts. In other words, how important is it for decision makers to understand an algorithm and are they influenced by the reputation of the
developer/provider of the algorithm. The study will consist of four treatments (full factorial design of high/low algorithm complexity and high/low developer/provider reputation) in which participants will be asked to choose between making an own forecast or following the algorithm. The monetary incentive will be based on forecast accuracy.
Intervention
In the control treatment, participants will be presented with a algorithm of low complexity and low reputation. Participants of the second treatment will be presented with an algorithm of high complexity and still low reputation. In treatments 3 and 4, reputation of the algorithm will be higher in both states of complexity.
Experimental design
Four treatments with full factorial design of low / high algorithm complexity and low / high reputation. In each treatment, participants are asked to forecast the demand for a product and place an order for that product for a total of 20 rounds. At the beginning of each round, they are presented with the values of three factors, of which only one is correlated with demand. Based on this information they choose a forecasting method (own forecast or algorithm). With this choice it is also determined how the order amount is being set for this round. Regardless of their choice, they continue by entering their own forecast . However, only if they previously chose their own forecast will this entry determine the order amount. In cases where they selected the algorithm, the algorithm will determine the order amount. At the end of each round, participants are given the results of their own forecast and the algorithm as well as actual demand and their corresponding profit from selling the product.
Design details
Cited by
No citation data available for this registration yet.