Bounded-Loss Private Prediction Markets Rafael Frongillo - PowerPoint PPT Presentation

Bounded-Loss Private Prediction Markets Rafael Frongillo University of Colorado, Boulder Neural Information Bo Waggoner Microsoft Research, NYC Processing Systems, December 2018 1 / 74

Prediction markets prediction p (0) learning (aggregation) algorithm future event ? time 2 / 74

Prediction markets Menu of purchases (gambles) learning (aggregation) algorithm future event ? time 3 / 74

Prediction markets purchase learning (aggregation) choice algorithm future event ? time 4 / 74

Prediction markets updated prediction p (1) learning (aggregation) algorithm future event ? time 5 / 74

Prediction markets updated prediction p (2) learning (aggregation) algorithm future event ? time 6 / 74

Prediction markets updated prediction p (t) learning (aggregation) algorithm future event ? time 7 / 74

Prediction markets observe event time 8 / 74

Prediction markets payments observe event time 9 / 74

Prior work Abernethy-F.-W. 2015 ( Neural Information Processing Systems ) ● 10 / 74

Prior work Abernethy-F.-W. 2015 ( Neural Information Processing Systems ) ● - differentially private prediction markets 11 / 74

Prior work Abernethy-F.-W. 2015 ( Neural Information Processing Systems ) ● - differentially private prediction markets - financial loss may not be bounded! 12 / 74

Prior work Abernethy-F.-W. 2015 ( Neural Information Processing Systems ) ● - differentially private prediction markets - financial loss may not be bounded! Cummings, Pennock, Wortman Vaughan 2016 ( EC ) ● - Impossibility: all private market makers have unbounded financial loss! 13 / 74

This paper Main result: a market construction with: differential privacy ● incentives to participate ● accuracy/fidelity of predictions ● Bounded worst-case financial loss ● 14 / 74

This paper Main result: a market construction with: differential privacy ( ε ) ● incentives to participate ( α ) ● accuracy/fidelity of predictions ( α , dimension d ) ● Bounded worst-case financial loss Õ( d / ε α ) ● Extensions (cf AFW’15): purchasing data, kernel methods…. 15 / 74

This paper Main result: a market construction with: differential privacy ( ε ) ● incentives to participate ( α ) ● accuracy/fidelity of predictions ( α , dimension d ) ● Bounded worst-case financial loss Õ( d / ε α ) ● Extensions (cf AFW’15): purchasing data, kernel methods…. Thanks! 74 / 74