Back to our work
Applied ML

Optimising real-time bidding on The Trade Desk

A bid optimisation system for advertising campaigns on The Trade Desk: event data pipelines, simulation-tested algorithms and services that adjust bids across many dimensions every hour.

Client
Jun Group
Period
Sep 2020 - Mar 2021
Our scope
Data, algorithms & bidding services
Engagement
Delivered by the aiLabs founders
Simulation and live bidding for Jun Group Campaign data supports simulation and hourly bid updates. Live results feed back into the campaign data. CTR was 30% higher than the control campaign in the client’s final live test. Campaign data Simulation Hourly bids Live results +30% CTR
Explanatory diagram illustrating the engineering approach.

Key results

  • In the client’s final acceptance test with live traffic, the click-through rate (CTR) was 30% higher than in the control campaign, with a noticeably lower cost per thousand impressions (CPM) and cost per click (CPC).
  • The client accepted the solution as meeting all requirements after dozens of tests. It was launched in production.
  • A Monte Carlo simulator for comparing bidding strategies before spending budget on live tests.

The challenge

Jun Group runs advertising campaigns on The Trade Desk and wanted better campaign KPIs through dimensional bidding. The platform accepts bid adjustments as bid lists about once an hour, clicks are rare (CTR around 0.1%), only won auctions are visible, and the platform’s own optimisation runs at the same time.

Our contribution

  • Analysed the platform’s event data and APIs, documented data gaps and designed the bid optimisation system.
  • Built Go ETL services that load raw impression, click, conversion and video event data from Amazon S3, and reference data from the platform API, into Google BigQuery, with PostgreSQL and ClickHouse options.
  • Replaced streaming inserts, which lost data at high volume, with Avro and CSV files loaded through Google Cloud Storage.
  • Developed bidding services that combine a bandit algorithm based on Thompson sampling with linear bidding proportional to CTR across up to 14 configurable dimensions, shifting from exploration to exploitation over the campaign period.
  • Added site blocking with a statistical threshold, base bid adjustment by hour of the week, budget pacing and max bid control, all applied through the platform API.
  • Built a Monte Carlo auction simulator with per-slice regression models, and used ANOVA to choose which dimensions to combine.

What the work delivered

The delivered system collected campaign data, tested strategies in simulation and adjusted live bids every hour. The client’s final live acceptance test confirmed that the solution met its requirements.

More engineering stories

View all work
Computer vision

Planogram compliance

As part of the client’s team, we built the 2D matching engine and backend. Cutting matching time for 80 products on one side of a fixture from 53 to 3-4 seconds made photo-based planogram checks practical on the shop floor.

Platform engineering

Platform modernisation

Cloud-independent infrastructure, event-driven services and a faster, cheaper data layer for an established learning analytics platform.

Computer vision

Computer vision platform

A computer vision platform that takes a project from data storage and annotation through training to live inference. Our team built two applications to demonstrate the full workflow.

Let’s build something useful

Bring us the
hard part.

A new AI capability, a complex data problem or a platform that needs to evolve. Tell us what you’re working on.

Discuss your project

Start a conversation

info@ai-labs.tech

aiLabs Innovations FZ-LLC
Compass Building, Al Shohada Road, Al Hamra Industrial Zone-FZ, Ras Al Khaimah
United Arab Emirates
Licence No. 47014246

For an initial conversation, share your goal, the current system and the challenge you want to solve.