| Popis: |
Machine Learning (ML) and Blockchains have been two major technology disruptions in the last decade. On the one extreme, Blockchain decentralizes decision making power to a crowd of anonymous participants. On the other extreme, ML centralizes decision making into uninterpretable algorithms. The first chapter uses ML as a tool to study behavioral biases in the labor markets. The second and third chapter deal with strategic interaction of market participant with Blockchain and ML platforms respectively. The first chapter, co‐authored with Prof Param Vir Singh and Prof Kannan Srinivasan, examines biases arising from attractive appearance. We show that Preference Bias contributes to an attractiveness gap of 0.52% per year, adding to a 2.4% gap over a 15‐year career. Belief Bias does not have a statistically significant contribution in our sample of 43,533 MBA graduates. This finding is important because Belief Bias, arising from evaluators group‐level priors on subjects job fit, can be overcome by providing rich performance information. But, Preference Bias, arising from evaluators taste for social, romantic or marital relationship with attractive subjects, can be harder to eliminate. We make use of ML based image morhping of subject appearance and page ranking of subject career milestones to construct a pseudo random experiment on observational data.The second chapter, co‐authored with Prof Manmohan Aseri, Prof Param Vir Singh and Prof Kannan Srinivasan, examines peer to peer payments on Blockchains. We show that upgrade to Bitcoin payment throughput is rolled back by tacit collusion among Bitcoin miners. We identify an intervention of banning miners beyond a maximum compute power to eliminate collusion. But, such an intervention makes payments less secure from double spend attacks. Thus owing to the dual threat, of collusion and double spend attacks, its untenable to offer a high througput payment ledger to users with widely different willingness to pay fees, bear delay and risk attacks. We advocate miner collusion as a useful mechanism where a chunk of excess collusion revenue endogenously spill over into an investment into platform’s security. The second chapter examines Machine Learning (ML) pricing in housing market. These MLmodels are revised regularly using recent sample of sales. The recent sales are themselves confounded by previous version of the ML model. We theoretically show how this Feedback Loop creates a self fulfilling prophecy where ML over estimates its own prediction accuracy and market participants over rely on ML predictions. We formulate size of resulting pricing bias. We identify conditions on ML and market characteristics such that participants are worse off after introduction of ML. We use data from Zillow’s Zestimate to provide empirical evidence for necessary primitives of our theoretical model. |
