Song Yao
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Song Yao has a total of 53 co-authors. Here are the top co-authors by number of shared publications:
Carl F. Mela
Duke University - Fuqua School of Business
4 shared publications- Stephan Seiler
Imperial College Business School; Centre for Economic Policy Research
4 shared publications - Wenbo Wang
Hong Kong University of Science and Technology
4 shared publications Chong Bo Wang
3 shared publications- Raphael Thomadsen
Olin Business School, Washington University in St. Louis, St. Louis, Missouri 63130
3 shared publications Tat Chan
3 shared publications
Latest Publications
- Open Access
- 19 August 2026
- Published by Elsevier in SSRN Electronic Journal
Stranded Credentials: How A Skill-Signaling Market Absorbed Generative AI
Generative AI can now perform many tasks that credentialing institutions count on to assess skill. During the AI era, do credentials retain their signaling value for subsequent performance? Mostly, yes. We audit the 2010-2026 archive of Kaggle, the largest data science competition platform, which ran two evaluation formats concurrently: upload-competitions, which directly score entrants' predictions computed on published data, and code-competitions, which score predictions by executing entrants' code on hidden data. Across 444,698 participations, competition medals predict subsequent leaderboard performance almost entirely in the first year after being earned, in both formats. Fresh medals retained most of their signaling value through the AI transition; credential stocks are only as informative as their replenishment. Although upload-competition medal stocks lost 82% of their informativeness, institutional stranding explains half to three quarters of the loss: upload-competitions had exited for reasons predating AI, and their frozen medal stock aged out under the pre-existing decay pattern. Old upload-competition medals look more valuable only in isolation, by proxying for the rest of the holder's record (e.g., experience). The measured changes are institutional rather than personal: an AI-like working style predicts performance similarly in both formats. The platform's official credential tiers, based on lifetime medal counts, discard 13-16% of the medals' information; an index weighting recent medals more heavily, built on pre-AI-era data alone, outperforms the official tiers in predicting AI-era performance. In conclusion, credentials are informative, perishable, institution-bound, and interdependent; sustaining their value under AI is a high-stakes, socioeconomic problem of institutional design.
- 19 August 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessStranded Credentials: How A Skill-Signaling Market Absorbed Generative AI
Generative AI can now perform many tasks that credentialing institutions count on to assess skill. During the AI era, do credentials retain their signaling value for subsequent performance? Mostly, yes. We audit the 2010-2026 archive of Kaggle, the largest data science competition platform, which ran two evaluation formats concurrently: upload-competitions, which directly score entrants' predictions computed on published data, and code-competitions, which score predictions by executing entrants' code on hidden data. Across 444,698 participations, competition medals predict subsequent leaderboard performance almost entirely in the first year after being earned, in both formats. Fresh medals retained most of their signaling value through the AI transition; credential stocks are only as informative as their replenishment. Although upload-competition medal stocks lost 82% of their informativeness, institutional stranding explains half to three quarters of the loss: upload-competitions had exited for reasons predating AI, and their frozen medal stock aged out under the pre-existing decay pattern. Old upload-competition medals look more valuable only in isolation, by proxying for the rest of the holder's record (e.g., experience). The measured changes are institutional rather than personal: an AI-like working style predicts performance similarly in both formats. The platform's official credential tiers, based on lifetime medal counts, discard 13-16% of the medals' information; an index weighting recent medals more heavily, built on pre-AI-era data alone, outperforms the official tiers in predicting AI-era performance. In conclusion, credentials are informative, perishable, institution-bound, and interdependent; sustaining their value under AI is a high-stakes, socioeconomic problem of institutional design.
- Open Access
- 29 July 2026
- Published by Elsevier in SSRN Electronic Journal
Equilibrium in the Sharing Economy: Field Experiments on Cross-Side Effects
Cross-side network effects are fundamental to sharing platforms, creating feedback loops that amplify initial shocks and shape final equilibrium. We quantify these effects in two separate field experiments-incorporating encouragement and switchback designs-on a ride-sharing platform. By instrumenting supply with driver subsidies and demand with rider coupons, we identify a significant asymmetry in network multipliers. A 1% initial increase in active drivers generates a 3.78-6.31% growth in equilibrium demand, nearly double the equilibrium supply growth. This reveals a counterintuitive result: despite increased competition, larger supply will enhance the use of individual drivers through improved matching efficiency. Conversely, demand-side shocks yield significantly smaller gains. Using a "residual filtering" strategy, we identify the effects of policy interventions. The results show that afternoon supply subsidies effectively stimulate both supply and demand through the evening. Counterfactual analysis suggests that a targeted $3 afternoon driver subsidy maximizes total platform revenue and profit. These results highlight the critical role of supply in driving platform growth and suggest that supply-side subsidies are an effective lever for scaling two-sided ecosystems.
- 29 July 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessEquilibrium in the Sharing Economy: Field Experiments on Cross-Side Effects
Cross-side network effects are fundamental to sharing platforms, creating feedback loops that amplify initial shocks and shape final equilibrium. We quantify these effects in two separate field experiments-incorporating encouragement and switchback designs-on a ride-sharing platform. By instrumenting supply with driver subsidies and demand with rider coupons, we identify a significant asymmetry in network multipliers. A 1% initial increase in active drivers generates a 3.78-6.31% growth in equilibrium demand, nearly double the equilibrium supply growth. This reveals a counterintuitive result: despite increased competition, larger supply will enhance the use of individual drivers through improved matching efficiency. Conversely, demand-side shocks yield significantly smaller gains. Using a "residual filtering" strategy, we identify the effects of policy interventions. The results show that afternoon supply subsidies effectively stimulate both supply and demand through the evening. Counterfactual analysis suggests that a targeted $3 afternoon driver subsidy maximizes total platform revenue and profit. These results highlight the critical role of supply in driving platform growth and suggest that supply-side subsidies are an effective lever for scaling two-sided ecosystems.
- Open Access
- 29 July 2026
- Published by Elsevier in SSRN Electronic Journal
When the Scaffold Stays On: AI, Practice Style, and Screening in Elite Skill Formation
Generative AI raises short-term productivity by completing tasks that learners would otherwise practice on their own. Whether this exchange erodes frontier skill depends on the mode of use: substitute-users let AI stand in for deliberate practice and fail to develop skill, while complement-users use it to accelerate skill development. For institutions that train and certify talent, the design question is not whether to allow AI but how to govern the mode of its use. We ask whether AI-prohibited evaluation gates can separate the two modes. In elite competitive programming, the International Collegiate Programming Contest (ICPC) and the International Olympiad in Informatics (IOI) prohibit AI under in-person proctoring, with qualification-round entry, whereas Codeforces (CF) practice is unproctored and open to all. From CF submission histories we build an AI-prompt signature, more first-attempt acceptances, fewer attempts, fewer debugging retries, consistent with AI-assisted practice. CF practice has shifted toward this signature across entry cohorts spanning two AI rollouts. In CF contests, a stronger signature predicts smaller rating gains for users with no ICPC-IOI affiliation, but not for those who qualified. Inside the AI-prohibited ICPC environment, a shift toward AI-style practice predicts higher non-AI-aided scores for AI-era entrants. The same signature carries opposite signs across the two environments, exactly the pattern a type-separating gate predicts. The message is constructive: AI-style practice is compatible with frontier skill; the erosion risk links to the substitute mode; and that mode is separable by gates standard at credential boundaries, from medical and legal boards to professional certification.
- 29 July 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessWhen the Scaffold Stays On: AI, Practice Style, and Screening in Elite Skill Formation
Generative AI raises short-term productivity by completing tasks that learners would otherwise practice on their own. Whether this exchange erodes frontier skill depends on the mode of use: substitute-users let AI stand in for deliberate practice and fail to develop skill, while complement-users use it to accelerate skill development. For institutions that train and certify talent, the design question is not whether to allow AI but how to govern the mode of its use. We ask whether AI-prohibited evaluation gates can separate the two modes. In elite competitive programming, the International Collegiate Programming Contest (ICPC) and the International Olympiad in Informatics (IOI) prohibit AI under in-person proctoring, with qualification-round entry, whereas Codeforces (CF) practice is unproctored and open to all. From CF submission histories we build an AI-prompt signature, more first-attempt acceptances, fewer attempts, fewer debugging retries, consistent with AI-assisted practice. CF practice has shifted toward this signature across entry cohorts spanning two AI rollouts. In CF contests, a stronger signature predicts smaller rating gains for users with no ICPC-IOI affiliation, but not for those who qualified. Inside the AI-prohibited ICPC environment, a shift toward AI-style practice predicts higher non-AI-aided scores for AI-era entrants. The same signature carries opposite signs across the two environments, exactly the pattern a type-separating gate predicts. The message is constructive: AI-style practice is compatible with frontier skill; the erosion risk links to the substitute mode; and that mode is separable by gates standard at credential boundaries, from medical and legal boards to professional certification.
- Open Access
- 29 July 2026
- Published by Elsevier in SSRN Electronic Journal
Unintended Consequences of Recommender System Interventions: Evidence from a Field Experiment
Platform content interventions in recommendation systems are typically evaluated as static "nudges," ignoring that the systems adaptively learn from the resulting user behavior. We investigate this dynamic through a large-scale field experiment on a short-video platform. The experiment involves a "sleep reminder" campaign designed to reduce late-night usage. Paradoxically, the intervention increased late-night engagement by 14.75% and overall platform usage by 2.18%, and the effects persisted for weeks even after the experiment. We explain this through a forced-exploration mechanism, showing that by revealing high latent demand for the promoted content, the intervention triggers a recommendation policy update that routine user behavior would not produce. The data generated by the intervention induced the algorithm to update its post-campaign policy, reinforcing the very engagement loops the campaign aimed to mitigate. Our findings demonstrate that user-facing interventions can effectively retrain the underlying algorithm, triggering durable, system-wide shifts in content distribution that challenge standard evaluation metrics in platform governance and social responsibility initiatives.
- 29 July 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessUnintended Consequences of Recommender System Interventions: Evidence from a Field Experiment
Platform content interventions in recommendation systems are typically evaluated as static "nudges," ignoring that the systems adaptively learn from the resulting user behavior. We investigate this dynamic through a large-scale field experiment on a short-video platform. The experiment involves a "sleep reminder" campaign designed to reduce late-night usage. Paradoxically, the intervention increased late-night engagement by 14.75% and overall platform usage by 2.18%, and the effects persisted for weeks even after the experiment. We explain this through a forced-exploration mechanism, showing that by revealing high latent demand for the promoted content, the intervention triggers a recommendation policy update that routine user behavior would not produce. The data generated by the intervention induced the algorithm to update its post-campaign policy, reinforcing the very engagement loops the campaign aimed to mitigate. Our findings demonstrate that user-facing interventions can effectively retrain the underlying algorithm, triggering durable, system-wide shifts in content distribution that challenge standard evaluation metrics in platform governance and social responsibility initiatives.
- Open Access
- 24 July 2026
- Published by Elsevier in SSRN Electronic Journal
When the Scaffold Stays On: AI, Practice Style, and Screening in Elite Skill Formation
Generative AI raises short-term productivity by completing tasks that learners would otherwise practice on their own. Whether this exchange erodes frontier skill, the capability required for top-tail performance without AI, depends on the mode of use: substitute-users let AI stand in for deliberate practice and fail to develop skill, while complement-users use it to accelerate skill development. For institutions that train and certify talent, the design question is therefore not whether to allow AI but how to govern the mode of its use. We study one governance mechanism, the AI-prohibited evaluation gate, and ask whether it can separate the two modes. Elite competitive programming provides a test case. The International Collegiate Programming Contest (ICPC) and the International Olympiad in Informatics (IOI) prohibit AI under in-person proctoring and admit entrants through qualification rounds, whereas online Codeforces (CF) practice is unproctored and open to all. From CF submission histories we build an AI-prompt signature, marked by more first-attempt acceptances and fewer attempts and debugging retries, consistent with AI-assisted practice. Three findings follow. CF practice has shifted toward this signature across entry cohorts spanning two AI rollouts. In open CF contests, a stronger signature predicts smaller rating gains for users with no ICPC or IOI affiliation, but not for users who qualified for the AI-prohibited contests. Inside the AI-prohibited ICPC environment, a shift toward AI-style practice predicts higher non-AI-aided scores for AI-era entrants. The same AI-prompt signature carries opposite signs depending on whether the contest screens its entrants through an AI-prohibited qualification gate, exactly the pattern a type-separating gate predicts. The message is constructive: AI-style practice is compatible with frontier skill; the skill erosion risk links to the substitute mode of use; and that mode is separable by the kind of gates already standard at credential boundaries, from medical and legal boards to professional certification.
- 24 July 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessWhen the Scaffold Stays On: AI, Practice Style, and Screening in Elite Skill Formation
Generative AI raises short-term productivity by completing tasks that learners would otherwise practice on their own. Whether this exchange erodes frontier skill, the capability required for top-tail performance without AI, depends on the mode of use: substitute-users let AI stand in for deliberate practice and fail to develop skill, while complement-users use it to accelerate skill development. For institutions that train and certify talent, the design question is therefore not whether to allow AI but how to govern the mode of its use. We study one governance mechanism, the AI-prohibited evaluation gate, and ask whether it can separate the two modes. Elite competitive programming provides a test case. The International Collegiate Programming Contest (ICPC) and the International Olympiad in Informatics (IOI) prohibit AI under in-person proctoring and admit entrants through qualification rounds, whereas online Codeforces (CF) practice is unproctored and open to all. From CF submission histories we build an AI-prompt signature, marked by more first-attempt acceptances and fewer attempts and debugging retries, consistent with AI-assisted practice. Three findings follow. CF practice has shifted toward this signature across entry cohorts spanning two AI rollouts. In open CF contests, a stronger signature predicts smaller rating gains for users with no ICPC or IOI affiliation, but not for users who qualified for the AI-prohibited contests. Inside the AI-prohibited ICPC environment, a shift toward AI-style practice predicts higher non-AI-aided scores for AI-era entrants. The same AI-prompt signature carries opposite signs depending on whether the contest screens its entrants through an AI-prohibited qualification gate, exactly the pattern a type-separating gate predicts. The message is constructive: AI-style practice is compatible with frontier skill; the skill erosion risk links to the substitute mode of use; and that mode is separable by the kind of gates already standard at credential boundaries, from medical and legal boards to professional certification.
- 1 March 2025
- Published by Institute for Operations Research and the Management Sciences (INFORMS) in Information Systems Research
- Vol. 36 (1) , 534-551
Customer Acquisition via Explainable Deep Reinforcement Learning
Effective customer acquisition is crucial for digital platforms, with sequential targeting ensuring that marketing messages are both timely and relevant. The proposed deep recurrent Q-network with attention (DRQN-attention) model enhances this process by optimizing long-term rewards and increasing decision-making transparency. Tested with a data set from a digital bank, the DRQN-attention model has proven to enhance clarity in decision making and outperform traditional methods in boosting long-term rewards. Its attention mechanism acts as a strategic tool for forward planning, pinpointing crucial ad marketing channels that are likely to engage and convert prospects. This capability enables marketers to understand the dynamic targeting strategies of the proposed model that align with customer profiles, dynamic behaviors, and the seasonality of the markets, thereby boosting confidence and effectiveness in their customer acquisition strategies.
- 1 March 2025
- Published by Institute for Operations Research and the Management Sciences (INFORMS) in Information Systems Research
- Vol. 36 (1) , 534-551
Customer Acquisition via Explainable Deep Reinforcement Learning
Effective customer acquisition is crucial for digital platforms, with sequential targeting ensuring that marketing messages are both timely and relevant. The proposed deep recurrent Q-network with attention (DRQN-attention) model enhances this process by optimizing long-term rewards and increasing decision-making transparency. Tested with a data set from a digital bank, the DRQN-attention model has proven to enhance clarity in decision making and outperform traditional methods in boosting long-term rewards. Its attention mechanism acts as a strategic tool for forward planning, pinpointing crucial ad marketing channels that are likely to engage and convert prospects. This capability enables marketers to understand the dynamic targeting strategies of the proposed model that align with customer profiles, dynamic behaviors, and the seasonality of the markets, thereby boosting confidence and effectiveness in their customer acquisition strategies.
- 1 May 2024
- Published by Institute for Operations Research and the Management Sciences (INFORMS) in Management Science
- Vol. 70 (5) , 3302-3318
The Impact of Government Interventions on COVID-19 Spread and Consumer Spending
We examine the impact of government interventions on the spread of COVID-19 and consumer spending. We do this by first estimating models of COVID-19 spread, consumer spending, and social distancing in the United States during the early stages of the COVID-19 pandemic. Social distancing has a large effect on reducing COVID-19 spread and is responsive to national and local case numbers. Nonmask government interventions reduce COVID-19 spread, whereas the effectiveness of mask mandates is much smaller and statistically insignificant. Mask mandates tend to increase social distancing, as do nonmask governmental restrictions as a whole. Social distancing hurts spending in the absence of a mask mandate but has a negligible effect on spending if there is a mask mandate. Mask mandates have a direct effect of increasing spending in counties with high levels of social distancing while reducing spending in counties with low levels of social distancing. We use these three estimated models to calculate the effect of mask mandates and other governmental interventions on COVID-19 cases, deaths, and consumer spending. Implemented mask mandates decreased COVID-19 cases by a statistically insignificant 774,000 cases, saving 28,000 lives, over a four-month period, but led to $76B–$155B of additional consumer spending. Other nonmask governmental interventions that were implemented reduced the number of COVID-19 cases by 34M, saving 1,230,000 lives, while reducing consumer spending by approximately $470B–$703B over our 4-month period of the study. Thus, these restrictions were cost effective as long as one values each saved life at $387,000–$608,000 or more. This paper was accepted by Matt Shum, marketing. Supplemental Material: The data is available at https://doi.org/10.1287/mnsc.2023.4853 .
- 1 May 2024
- Published by Institute for Operations Research and the Management Sciences (INFORMS) in Management Science
- Vol. 70 (5) , 3302-3318
The Impact of Government Interventions on COVID-19 Spread and Consumer Spending
We examine the impact of government interventions on the spread of COVID-19 and consumer spending. We do this by first estimating models of COVID-19 spread, consumer spending, and social distancing in the United States during the early stages of the COVID-19 pandemic. Social distancing has a large effect on reducing COVID-19 spread and is responsive to national and local case numbers. Nonmask government interventions reduce COVID-19 spread, whereas the effectiveness of mask mandates is much smaller and statistically insignificant. Mask mandates tend to increase social distancing, as do nonmask governmental restrictions as a whole. Social distancing hurts spending in the absence of a mask mandate but has a negligible effect on spending if there is a mask mandate. Mask mandates have a direct effect of increasing spending in counties with high levels of social distancing while reducing spending in counties with low levels of social distancing. We use these three estimated models to calculate the effect of mask mandates and other governmental interventions on COVID-19 cases, deaths, and consumer spending. Implemented mask mandates decreased COVID-19 cases by a statistically insignificant 774,000 cases, saving 28,000 lives, over a four-month period, but led to $76B–$155B of additional consumer spending. Other nonmask governmental interventions that were implemented reduced the number of COVID-19 cases by 34M, saving 1,230,000 lives, while reducing consumer spending by approximately $470B–$703B over our 4-month period of the study. Thus, these restrictions were cost effective as long as one values each saved life at $387,000–$608,000 or more. This paper was accepted by Matt Shum, marketing. Supplemental Material: The data is available at https://doi.org/10.1287/mnsc.2023.4853 .
- Open Access
- 1 January 2024
- Published by Elsevier in SSRN Electronic Journal
Customer Acquisition Via Explainable Deep Reinforcement Learning
Effective customer acquisition heavily hinges on sequential targeting to ensure appropriate marketing messages reach customers. Sequential targeting could guide customers through the acquisition process and thus optimize long-term revenue for the firm. Towards this goal, Reinforcement Learning (RL) has demonstrated great potential in facilitating sequential targeting during user acquisition. However, decisions made by RL during this process often lack explainability. We introduce the DRQN-Attention model, which optimizes the long-term reward of sequential targeting while enhancing the explainability of the decisions. The key idea of the proposed model is to revise Q-Learning by adding an attention mechanism to create a bottleneck, forcing the model to focus on features of the next ad exposure that will lead to optimal long-term rewards. We estimate our model using a comprehensive dataset from a digital bank. The empirical results show the proposed model is explainable and also outperforms state-of-the-art methods in terms of long-term revenue optimization. Specifically, the attention mechanism within the model functions as forward planning. The forward planning can spot those features in the next ad exposure that are more likely to lead to the optimal outcome. We further demonstrate how the model makes targeting decisions of advertising channel choices by showing that the model can: 1) learn optimal ad channels to target customers from different industries, 2) adjust advertising channels in response to dynamic customer behaviors, and 3) learn the seasonality of the customer’s industry and calibrate ad channel correspondingly.
- 1 January 2024
- Published by Elsevier in SSRN Electronic Journal
Open AccessCustomer Acquisition Via Explainable Deep Reinforcement Learning
Effective customer acquisition heavily hinges on sequential targeting to ensure appropriate marketing messages reach customers. Sequential targeting could guide customers through the acquisition process and thus optimize long-term revenue for the firm. Towards this goal, Reinforcement Learning (RL) has demonstrated great potential in facilitating sequential targeting during user acquisition. However, decisions made by RL during this process often lack explainability. We introduce the DRQN-Attention model, which optimizes the long-term reward of sequential targeting while enhancing the explainability of the decisions. The key idea of the proposed model is to revise Q-Learning by adding an attention mechanism to create a bottleneck, forcing the model to focus on features of the next ad exposure that will lead to optimal long-term rewards. We estimate our model using a comprehensive dataset from a digital bank. The empirical results show the proposed model is explainable and also outperforms state-of-the-art methods in terms of long-term revenue optimization. Specifically, the attention mechanism within the model functions as forward planning. The forward planning can spot those features in the next ad exposure that are more likely to lead to the optimal outcome. We further demonstrate how the model makes targeting decisions of advertising channel choices by showing that the model can: 1) learn optimal ad channels to target customers from different industries, 2) adjust advertising channels in response to dynamic customer behaviors, and 3) learn the seasonality of the customer’s industry and calibrate ad channel correspondingly.
- Open Access
- 1 January 2024
- Published by Elsevier in SSRN Electronic Journal
The Effect of the Order of Incentive Framing on Performance
Companies often provide incentives to influence employee performance or customer engagement. In this research, we investigate how companies structure their incentives so that they increase the performance and engagement of their employees and customers without increasing the rewards. In particular, we examine situations where companies change the way they frame and order their goals to incentivize their employees and customers. Across a lab and a field experiment, when rewards do not change, we find that performance increases when goals are perceived as becoming less challenging over time. In the lab experiment, consumers play a video game, and we track the influence of altering the way customers’ goals are framed and ordered over time on their scores. In the field experiment, we partnered with a ride-sharing company to examine the influence of the way that employees’ goals are structured on the number of rides taken. Subsequently, our research finds that performance can be enhanced when goals are structured such that they are perceived as getting less challenging over time, without needing to spend more money on additional rewards.
- 1 January 2024
- Published by Elsevier in SSRN Electronic Journal
Open AccessThe Effect of the Order of Incentive Framing on Performance
Companies often provide incentives to influence employee performance or customer engagement. In this research, we investigate how companies structure their incentives so that they increase the performance and engagement of their employees and customers without increasing the rewards. In particular, we examine situations where companies change the way they frame and order their goals to incentivize their employees and customers. Across a lab and a field experiment, when rewards do not change, we find that performance increases when goals are perceived as becoming less challenging over time. In the lab experiment, consumers play a video game, and we track the influence of altering the way customers’ goals are framed and ordered over time on their scores. In the field experiment, we partnered with a ride-sharing company to examine the influence of the way that employees’ goals are structured on the number of rides taken. Subsequently, our research finds that performance can be enhanced when goals are structured such that they are perceived as getting less challenging over time, without needing to spend more money on additional rewards.
- 1 September 2023
- Published by Institute for Operations Research and the Management Sciences (INFORMS) in Management Science
- Vol. 69 (9) , 5405-5417
The Impact of a Supreme Court Decision on the Preferences of Americans Regarding Abortion Policy
- 1 September 2023
- Published by Institute for Operations Research and the Management Sciences (INFORMS) in Management Science
- Vol. 69 (9) , 5405-5417
The Impact of a Supreme Court Decision on the Preferences of Americans Regarding Abortion Policy