A Parimutuel-like Mechanism for Information Aggregation: Forecasting Inside Intel∗

We evaluate the performance of an information aggregation mechanism (IAM) implemented inside Intel to forecast unit sales for the company. Developed, refined and tested in the laboratory using experimental methods, the IAM is constructed from institutional features that improve performance and overcome barriers to successful applications of other information aggregation methods. Its implementation at Intel provides a testbed for evaluating this new form of IAM’s performance in a complex field environment. In contrast to prediction markets, which provide only a point forecast of future sales, the IAM characterizes the full distribution of participants’ aggregated beliefs allowing a more detailed evaluation of its performance. We show this predictive distribution very closely matches the distribution over outcomes at short horizons while slightly underweighting low-probability realizations of unit sales at long horizons. Compared to Intel’s “official forecast,” the IAM forecasts perform well overall, even though they predate the official forecasts. The forecast improvements are most prominent at short forecast horizons and in direct distribution channels, where the effective aggregation of individually-held information drives the IAM to be more accurate than the official forecast over 75% of the time. ∗The financial support of the Lee Center for Advanced Networking, the Gordon and Betty Moore Foundation, and the Laboratory of Experimental Economics and Political Science is gratefully acknowledged. We thank Dan Zhou for providing excellent research assistance, and Chew Soo Hong, Erik Snowberg, Allan Timmermann, Michael Waldman, and seminar audiences at Arizona State, EIEF Roma, MIT, Rice, UCSD, Universidad Carlos III Madrid, IIOC 2013, and the Econometric Society NASM 2013 for helpful comments.