A rule for choosing when to sell/buy an asset when facing a sequence of prices
This Research Card explains a rule for choosing when to sell/buy an asset when facing a sequence of prices. The key feature is that the algorithm has access to “predictions” on the optimal price. The objectives are to take “safe” decisions, in the sense that if predictions are wrong the loss is not too much compared to the typical decision without predictions, yet “agressive enough” so that accurate predictions can be strongly leveraged. There is a tradeoff to analyse between those two antagonist objectives.
- Title: Pareto-Optimality, Smoothness, and Stochasticity in Learning-Augmented One-Max-Search
- Authors: Ziyad Benomar, Lorenzo Croissant, Vianney Perchet, Spyros Angelopoulos
- Team: R&D — CAIL — Fairplay
- Revue: ICML 2025
- Status: Accepted
- Category: Bidding
Pareto-Optimality, Smoothness, and Stochasticity in Learning-Augmented One-Max-Search
One-max search is a classic problem in online decision-making, in which a trader acts on a sequence of revealed prices…
arxiv.org
Why did we work on this topic? What problem did we want to solve?
Ad Tech systems use machine learning to predict user value or auction price. But small mistakes can flip a “bid” decision to “no bid,” causing missed chances or extra cost. We wanted to make a rule that:
- Uses accurate predictions fully,
- Protects you in the worst price sequences, and
- Lowers its performance smoothly when predictions get worse, so you avoid sudden drops in spending or ROI.
What did we find? What did we achieve?
- We made a simple rule that knows when to bid high or stay low.
- We added a smoothness knob so that if our guesses get worse, our results go down slowly, not suddenly.
- We showed that this rule still works even when prices and forecasts jump around randomly, and we tell you roughly how much money you can make.
How did we proceed?
This is an upstream, fondamental research paper, motivated by a concrete use case. The first step is one of the most difficult, and consists in defining the appropriate mathematical model… and the second one to analyse it using all our skills and techniques !
What is the originality here?
The main difference with the state of the art is that we assume the predictions to be stochastic, or random, and more or less correlated to the true value of the optimum. We are able to quantify the complexity of the problem at hand using those new concepts.
Check all our research cards 👇
https://medium.com/criteo-engineering/research-cards/home
