Description

We study whether autonomous pricing algorithms sustain supracompetitive prices in a spatially differentiated market, using Berlin's inner-Ringbahn grocery retail sector (≈493 stores, three chain types: discount, standard, bio) as an empirical testbed. Demand follows a spatial multinomial logit with vertical differentiation via a social-status-based willingness-to-pay proxy. Store-level Q-learning agents price tactically within a spatial logit demand system; a subset of specifications adds an LLM-based chain-level "CEO" layer that sets periodic price envelopes constraining the tactical agents below. Collusion is measured via the Calvano et al. (2020) Δ index against calibrated Bertrand-Nash and joint-monopoly benchmarks.

Q-learning agents alone sustain only weak, spatially localized collusion — Δ stays close to the noncooperative benchmark, consistent with dilution of strategic interaction across a large number of largely independent store-level state spaces. Adding an LLM CEO layer changes this qualitatively: pricing converges to near-perfect collusion, and this holds regardless of whether CEOs can communicate with rival chains. Analysis of the envelope mechanism shows coordination operates primarily through focal-price placement rather than band-width adjustment. The decisive factor driving collusion is not inter-agent communication but whether CEOs are given explicit knowledge of the Bertrand-Nash and joint-monopoly reference prices — this benchmark information alone is sufficient to induce near-complete supracompetitive coordination, even absent any communication channel, with implications for algorithmic antitrust enforcement.