A fixed discount cannot tell the difference between an event that is about to sell out and one with seats still to fill. The price should reflect the situation, not just the calendar.
The problem with fixed discounts
Last-minute tickets were priced using static discount bands. That could give away margin when an event was already selling well, while another event still reached opening time with seats that could no longer generate revenue.
The client needed prices to react to the actual booking situation without giving up control or creating unpredictable changes for customers.
A pricing engine that learned from each outcome
We built a dynamic pricing engine connected to the client’s existing booking system through an API, a controlled connection that lets software exchange data. It wrote each decision and outcome to the client’s Postgres database. For every booking request, the engine considered remaining tickets, event capacity, time until the event and recent sales. It compared only the prices inside the client’s minimum and maximum limits using expected revenue: ticket price multiplied by the probability of sale. A short-lived price lock then kept the selected price stable while the customer booked.
At its core, the engine used Bayesian updating, a structured way to revise an estimate as new evidence arrives. It began with an estimate of how likely each price was to result in a sale, then strengthened or weakened those estimates as real sale and no-sale outcomes arrived. Thompson sampling balanced using the prices supported by stronger evidence with a controlled chance to test another allowed price. This helped the system learn without uncontrolled price changes.
Every price and outcome was recorded in Postgres. A dashboard showed what the system was learning, and a controlled A/B comparison tested the new engine against the previous pricing rules. This gave the client evidence to decide whether the system produced more revenue from the same ticket inventory before expanding its use.
The engine estimated demand at each allowed price, chose the option with the best expected revenue, then learned from whether the customer bought.
Every sale or no-sale refined the estimate used for future decisions. Prices stayed within the client’s limits, and a short-lived lock kept the displayed price stable during booking.
What improved
- Revenue-aware protects margin or responds when seats risk going unsold
- Controlled every price stays inside business limits and remains stable at checkout
- Provable the new engine is compared directly with the existing rules
The pointDynamic pricing is useful only when the business can control it, customers experience it consistently and the commercial result can be proved.