Toward Algorithm Market Success: Code, Data, and (Pre-trained) Models

Artificial intelligence (AI) and machine learning have become an important componentof many business applications. However, firms still face the question of how to acquirethe capabilities they need to develop machine learning algorithms. Traditionally, thereare two ways to conduct machine learning projects: outsourcing to vendors or developingin-house teams. This project studies a third channel: algorithm markets.Algorithm markets have existed in two forms: seller-oriented and buyer-oriented. Inseller-oriented markets, developers post their algorithms/models for customers to adopt.In buyer-oriented markets, customers post their problems and data for developers totackle. These markets have traditionally been studied within the e-commerce/app storeand crowdsourcing literature. However, algorithms are different from traditionalcommodities or crowdsourced services. They usually need to be customized more thancommodities and have more constraints than regular outsourced services. To effectivelymanage algorithm markets, it is necessary to understand their unique mechanisms.This project plans to investigate three unique aspects of algorithm markets: code, data,and (pre-trained) models. Code and data are basic elements of machine learningsolutions, where code determines model structure and data determines modelparameters. Pre-trained models are a new form of solution that combines code and data.This project will examine how the code representation will change the behavior ofconsumers and how datasets will change the behavior of developers. Next, it will explorehow pre-trained models will affect the dynamics of both types of stakeholders.In this investigation, this project will employ machine learning and econometric analysismethods. Machine learning will capture the characteristics of code, data, and pre-trainedmodels, which are too complicated for humans to conceptualize and measure. Rigorouseconometric methods will be used to provide causal identification of relations betweenthe variables so that the findings can be implemented by market makers.Theoretically, the project will deepen our understanding of algorithm markets and pushforward studies on the stakeholders of machine learning projects. Practically, the projectcould provide insights to the management of algorithm markets. Given the explosivegrowth of artificial intelligence, algorithm trading may become a big business sector, andHong Kong can consider building algorithm markets that connect the modelingcapabilities and demand between developed and developing regions. To strengthen itsrole as an international trade center, algorithmic assets should not be ignored by HongKong. The findings of this project may contribute to Hong Kong’s long-term success inthis line of business.