Nicolas Eschenbaum

We introduce Multi-Agent Deep Hedging (MADH), a computational framework that extends deep reinforcement learning to markets with endogenous price formation. MADH embeds a differentiable market-clearing mechanism into the learning process, enabling decentralized agents to internalize their price impact via gradient ascent. We apply MADH to peer-to-peer electricity trading, benchmarking it against a centralized welfare-maximizing planner. Using synthetic data for heterogeneous prosumer communities, we demonstrate that decentralized agents autonomously learn sophisticated arbitrage strategies, such as capacity withholding. Crucially, we find that this strategic behavior generates positive externalities: while active traders reduce their own costs through price-awareness, their arbitrage smooths market prices, reducing costs for passive consumers. Furthermore, quantitative regret analysis confirms that MADH policies converge with low regret < 1.5%. These results establish MADH as a scalable tool for designing stable and efficient autonomous trading platforms.

This article examines the extent to which designated “gatekeepers” implement the provisions of the EU’s Digital Markets Act outside its territorial scope ("Brussels Effect"). Drawing on transparency reports, contractual documents, and informal communications, we reveal significant disparities in compliance strategies. Apple, Google, and Booking predominantly restrict their implementation to the EU or EEA, whereas Microsoft, Meta, and ByteDance extend certain measures to non-EU jurisdictions, notably Switzerland. Crucially, obligations subject to non-compliance proceedings by the European Commission are rarely extended beyond the EU, suggesting a strategic approach to territorial extensions of the DMA’s implementation. The article also uncovers a pattern of complex and sometimes contradictory communication by gatekeepers, raising questions about the transparency of the DMA’s implementation. These inconsistencies, coupled with selective extensions of specific data-related DMA provisions, point to a fragmented “Brussels Effect” of the law. The findings also imply that gatekeepers weigh the economic and strategic costs of compliance when deciding on territorial scope, and that the DMA’s global impact may depend on further coordination between regulators as well as more stringent enforcement.

This paper develops an analytical framework that captures a broad class of monopoly pricing problems, aiming to explain why Coasian dynamics emerge in some settings while Coasian failures arise in others. We introduce the notion of trading-up opportunities and show that they are the driving force behind Coasian dynamics. In particular, pricing dynamics do not emerge in the absence of trading-up opportunities—a Coasian failure. Instead, with trading-up opportunities, pricing dynamics arise until these opportunities are exhausted or the game ends. We show how our analysis generalizes to transitional games where one variety is only indirectly accessible.

This paper develops a formal framework to assess policies of learning algorithms in economic games. We investigate whether reinforcement-learning agents with collusive pricing policies can successfully extrapolate collusive behavior from training to the market. We find that in testing environments collusion consistently breaks down. Instead, we observe static Nash play. We then show that restricting algorithms’ strategy space can make algorithmic collusion robust, because it limits overfitting to rival strategies. Our findings suggest that policy-makers should focus on firm behavior aimed at coordinating algorithm design in order to make collusive policies robust

This paper presents experimental evidence on dynamic pricing in a large number of finite-horizon bilateral relationships, building on Hart and Tirole (1988). We examine four distinct treatments that vary the mode of trade and the seller’s commitment ability. We find that theory accurately predicts average prices across relationships but falls short of capturing the diversity of individual price trajectories. We also find that commitment has less bite than theory predicts, with sellers leaving significant rents to buyers and frequently committing to changing or oscillating prices. Our analysis suggests that theory explains behavior under renting better than under selling.

We analyze the vulnerability of decentralized autonomous organizations (DAOs) to speculative exploitation via their redemption mechanisms. Studying a game-theoretic model of repeated auctions for governance shares with speculators, we characterize the conditions under which—in equilibrium—an exploitative exit is guaranteed to occur, occurs in expectation, or never occurs. We evaluate four redemption mechanisms and extend our model to include atomic exits, time delays, and DAO spending strategies. Our results highlight an inherent tension in DAO design: mechanisms intended to protect members from majority attacks can inadvertently create opportunities for costly speculative exploitation. We highlight governance mechanisms that can be used to prevent speculation.