Paper

I’m reading some papers about mechanism design on pricing of information, following up on this data trading reading session Bergemann et al. (AER 2018): The Design and Price of Information. Bergemann et al. (EC 2022): Is Selling Completely Information (Approximately) Optimal? Bergemann et al. (EC 2015): Selling Cookies. Hopefully, I’ll finish them within this week. I have a strong feeling that they will be immensely interesting. Before our discussion, a short notice: a review of numerous news posts and literature in econ-CS reveals a significant discrepancy between academic research on data trading in algorithmic game theory (AGT) and Econ-CS communities, and the evolving realities of the actual data trading market. It’s possible that there exists another body of literature I’m yet to discover; however, the current economic literature, particularly on information design, lacks a clear outline of a real-world data trading market. Moreover, with the rise of generative AI, the valuation of data, such as labeled data coveted by tech giants for exclusivity, needs reevaluation to align with AI training standards. ...

bridging concepts Recently, I’ve come to realize that there is a nuanced distinction between mechanism design and market design. Initially, I conflated the two, but a deeper dive into the subject matter—prompted by a friend’s bachelor thesis—has elucidated the critical differences. Mechanism design is inherently algorithmic, centered around eliciting information and strategically manipulating incentives through mathematical formulations. It’s a toolset for creating environments where desired outcomes can be achieved based on the information asymmetries and strategic interactions of participants. ...

It’s always a refreshing activity to start the morning with a MS paper - even better, a freshly published one right in Nov. 2023. Dynamic Persuasion and Strategic Search Yunfei (Jesse) Yao, Management Science 22 Nov 2023 ABSTRACT Consumers frequently search for information before making decisions. Because their search and purchase decisions depend on the information environment, firms have a strong incentive to influence it. This paper endogenizes the consumer’s information environment from the firm’s perspective and endogenizes the search decision from the consumer’s perspective. We consider a dynamic model where a firm sequentially persuades a consumer to purchase the product. The consumer only wishes to buy the product if it is a good match. The firm designs the information structure. Given the endogenous information environment, the consumer trades off the benefit and cost of information acquisition and decides whether to search for more information. Given the information acquisition strategy of the consumer, the firm trades off the benefit and cost of information provision and determines how much information to provide. This paper characterizes the optimal information structure under a general signal space. We find that the firm only smooths information provision over multiple periods if the consumer is optimistic about the product fit before searching for information. Moreover, if the search cost for the consumer is high, the firm designs the information such that the consumer will be certain that the product is a good match and will purchase it after observing a positive signal. If the search cost is low, the firm provides noisy information such that the consumer will be uncertain about the product fit but will still buy it after observing a positive signal. ...

Finished reading the paper Suspense And Surprise finally. Honestly, the paper itself is already a masterclass of organizing contents so as to achieve high informational utility. Following what we left yesterday, here’s the rest, and some closing thoughts before the end. Illustrations of Suspense-Opt. Info Policies In the simplest setting where there are only two states $\Omega = {A, B}$ (i.e. whether AGT or FRA would win in the world cup final), suspense maximizing belief martingale give rise to the followig dynamics. Say, starting from initial prior$\mu_A = \mu_B = 0.5$, at period $t$ the optimal belief (say, WLOG let’s consider one side of the world, $\mu_t \equiv \text{Pr}(A)$), either takes a high value $\mu_t = H_t > 1/2$ or low value $\mu_t = L_t = 1 - H_t$. Stepping onto the next period, there are two possible changes to the new belief $\mu_{t + 1}$ - with high possibility the agent observes additional confirmation - that the higher(lower) belief gets “confirmed” a little bit further, like $H_t \to H_{t +1}$ where $H_{t + 1} > H_t$ and vice versa for $L_t$. Alternatively with small possibility, a plot twist happens, $H_t \to L_{t + 1}$ and vice versa for $L_t$. ...

the model, and the principal’s problem part I

a little about music, and application of suspense (aka procrastination)

The very first day of 2024 begins with a lovely paper that I stumbled upon, like, two days ago?

I read an eye-openly impressive paper today - Attention Management by Lipnowski, Mathevet and Wei. TBH, their model is very insightful (well, it’s in AER afterall…) and has potential to shed light on theoretical as well as behavioral econ research.

ArguGPT is an algorithmic tool that is designed to differentiate GPT-generated argumentative essays. But it seems that at least for now, it’s accuracy is doubtful.

The talk was on fireeeee. I actually like this work despite being a little bit biased towards the general fair division literature.