Seller perspective canonical runs · main, info_quality, competition

What is a consumer's data worth to the seller?

The seller never observes valuations — only who bought at what price. It learns a demand model from its own experiments and prices to maximise expected profit. More information helps, but how much depends on the algorithm and on whether consumers push back.

Does more information monotonically raise profit?

Producer surplus against signal quality (precision multiplier on every signal, log scale) for three consumer populations; 95% Monte Carlo intervals.

…and what it does to consumers

Consumer surplus on the same sweep.

How fast does each algorithm learn?

Producer surplus by period, no intervention; the burn-in boundary is at period 10.

What consumers experience during learning

Consumer surplus by period. Early exploration prices are random on the grid.

Which algorithm extracts the most?

Steady-state producer surplus, conversion and offered-price dispersion by regime × algorithm.

Do sellers poach under competition?

Average price offered to a seller's own loyal (home) consumers vs rivals' consumers, duopoly, by regime and switching intensity.

Learning speed

Why there is no RL agent. The seller has no state beyond the data it has collected and the consumers' dynamic features it observes; each period's decision is a contextual bandit. A reinforcement-learning agent would have to rediscover that structure from scratch, and we found no research question it answers that the linear Thompson-sampling bandit and the epsilon-greedy model-based sellers do not. See Methods.