Welfare analysis canonical runs · main, consumer_mix, robustness, competition

Does personalization create or destroy value?

Textbook first-degree price discrimination is efficient: everyone with valuation above cost is served. Real algorithms have noisy information, and real consumers refuse, delay and hide. Both push against the textbook.

Where does every regime × policy sit on the consumer–producer plane?

Steady-state consumer surplus vs producer surplus for all 90 main-experiment cells. The dashed line is the first-best frontier (CS + PS = first best).

Where does the deadweight loss come from?

Efficiency (total welfare ÷ first best) and share of consumers served, by regime.

How much do the conclusions depend on consumer behaviour?

Total welfare change (individualized boosting minus uniform) under six behavioural mixes; 95% intervals.

…and the consumer-surplus change

Same mixes, consumer surplus.

Does competition rescue consumers?

Average price faced and consumer surplus by number of sellers; solid = high switching (0.5), dotted = low (0.1).

Robustness: individualized boosting vs uniform, % change

One-at-a-time deviations from the baseline (10 seeds each). Rows are axes, columns are outcomes.

Welfare assumptions made visible. Total welfare is the unweighted sum of consumer and producer surplus (a dollar is a dollar whoever holds it); privacy effort is a real cost; strategic hold-outs and reference-price refusals are counted as lost trades, not as consumer errors; bounded-rational over-payment counts as negative consumer surplus. Distributional weights would change the ranking of policies — the Consumer page and the CS Gini exist so a reader can apply their own.