Research programme
Elias Sultan: six years of marketplace decision systems research
An account of how this work developed — what I was doing when each question appeared, what I got wrong, and how the answers ended up in software.
Written in the first person, because the research was not conducted at a distance. Every finding below cost something before it was understood.
2020 — 2021
The problem arrived before the research did.
I started selling on the Amazon marketplace in Spain in 2020. Within months I was doing what everyone does: configuring a repricing tool, setting minimum and maximum prices, and letting it work.
It worked exactly as designed, and that was the problem. Prices moved constantly, positions changed hands, and at the end of the month the margin was worse than the month before — with no single decision I could point to as the mistake.
The first genuine observation came from watching a single listing for a week. My system lowered a price. Within a short interval, the competing offer lowered too. Neither of us gained position. Both of us earned less.
That is not a market finding a price. That is two machines executing instructions that neither operator would have given if asked directly.
2022 — 2023
Building instruments before building answers.
To study the behaviour I had to record it, and nothing available recorded what I needed. Aggregated pricing feeds gave me a number per listing; what I needed was the offer as a buyer sees it — which seller, at what price, with which delivery promise, from which fulfilment channel, and who held the featured position.
So the first years of this work produced no conclusions at all. They produced instrumentation.
Once the record existed, competitors stopped being anonymous prices. They became identifiable actors that behaved consistently enough to be characterised: one reacted within a predictable interval, another followed a price down only so far, a third abandoned a position under pressure that a fourth would defend.
The models built in this period were the first that asked whether a price change was worth making, rather than only what the new price should be. Most of them were wrong. They were wrong in ways I could measure, which was the point.
2024 — 2025
Two findings that changed how the system works.
Reaction is not instantaneous. Measuring the interval between my action and a competitor's response, per competitor, over months, showed that the delay is real and stable enough to be treated as a characteristic. Between the two there is a window in which the position is held at a price the competitor has not matched. That window is a resource, and a system that ignores it pays permanently for something it could hold temporarily.
This is documented as RP-001.
The costs I was pricing against were not the costs I was paying. Reconciling settled marketplace transactions to the unit level, against the estimates used to configure minimum prices, showed a divergence that was not uniform — it concentrated in particular weight bands, in products with higher return rates, and in promotional participation. Some products had been trading below their true floor while every dashboard reported them profitable.
This is documented as RP-004.
2026
Where the work stands.
The findings consolidated into a framework in which price is the output of a decision rather than the decision itself: signal is interpreted as behaviour, commercial constraint is applied, and the system chooses to contest, hold, withdraw or recover — only then producing a number. That framework is published in full as White Paper 001.
Its implementation is SellerFlow, the platform that runs the operation this research is conducted on. The research library is at eliassultan.pro/research, and the longer account of the method is on the about page.
What comes next is stated plainly in the paper's own section on limitations: this is one business, in one marketplace, over six years. The framework should transfer. The parameters will not. Establishing which is which is the work ahead.
Enquiries
Questions about the research?
Available for interviews, technical commentary and collaboration with researchers and industry publications.
Contact