Undergraduate thesis · Western University
Algorithmic pricing
When rivals get along.
Two pricing agents learn to compete. I studied when they stop—and what an audit can catch.
Run the experimentA reduced browser experiment: 9 prices, discount 0.99, 500k steps. Results vary by seed. The thesis uses 15 prices and discount 0.95; its measured results are reported separately below.
01 / The question
What can a detector actually prove?
02 / What I built
I reproduced the Calvano et al. pricing benchmark across 1,000 learning sessions, then froze the policies and evaluated them against exact best responses. Supervisor: Dr. Apurva Narayan.
I built behavioral screens, calibrated them against designated honest markets, and constructed policies that know how those tests work. Changing the detector changes the room those policies have to operate in.
03 / What happened
The reproduced normalized profit gain was 0.852 against the published 0.849, with a 95% interval of [0.845, 0.859]. Designed policies retained 70.7% of the competitive-to-monopoly gain against the full screen battery and 99.3% against the original reaction audit alone.
04 / Where it stops
A widened audit reduced the strongest credible evasion in the enumerated grim/ladder families to 37.8%. A later exploratory tier-and-stick construction passed that widened audit at 99.3%, so 38% is not a universal cap. These are controlled markets, calibrated samples, and deliberately constructed evaders; the browser experiment is a smaller illustration of the learning mechanism.
One more? / 03