Private collaborative forecasting and benchmarking

Suppose a number of hospitals in a geographic area want to learn how their own heart-surgery unit is doing compared with the others in terms of mortality rates, subsequent complications, or any other quality metric. Similarly, a number of small businesses might want to use their recent point-of-sales data to cooperatively forecast future demand and thus make more informed decisions about inventory, capacity, employment, etc. These are simple examples of cooperative benchmarking and (respectively) forecasting that would benefit all participants as well as the public at large, as they would make it possible for participants to avail themselves of more precise and reliable data collected from many sources, to assess their own local performance in comparison to global trends, and to avoid many of the inefficiencies that currently arise because of having less information available for their decision-making. And yet, in spite of all these advantages, cooperative benchmarking and forecasting typically do not take place, because of the participants' unwillingness to share their information with others. Their reluctance to share is quite rational, and is due to fears of embarrassment, lawsuits, weakening their negotiating position (e.g., in case of over-capacity), revealing corporate performance and strategies, etc. The development and deployment of <i>private</i> benchmarking and forecasting technologies would allow such collaborations to take place without revealing any participant's data to the others, reaping the benefits of collaboration while avoiding the drawbacks. Moreover, this kind of technology would empower smaller organizations who could then cooperatively base their decisions on a much broader information base, in a way that is today restricted to only the largest corporations. This paper is a step towards this goal, as it gives protocols for forecasting and benchmarking that reveal to the participants the desired answers yet do not reveal to any participant any other participant's private data. We consider several forecasting methods, including linear regression and time series techniques such as moving average and exponential smoothing. One of the novel parts of this work, that further distinguishes it from previous work in secure multi-party computation, is that it involves floating point arithmetic, in particular it provides protocols to securely and efficiently perform division.