Welcome to the journal
Welcome to The Agentic Trading Research Journal. This is the first entry in a series about building AI systems that study markets, make decisions, and learn from what happens next. We are interested in the engineering as much as the outcome: how a candidate is discovered, what evidence supports an entry, whether it could actually be sold, and how to tell a useful lesson from a lucky result.
Each edition will take one research question and follow it through a method, the available evidence, its limits, and a next experiment. We will write about agent design, market data, execution, risk controls, and evaluation. A trading-service update may help explain an experiment, but this journal is a place to examine the ideas behind the work. We will distinguish funded trades from paper observations and avoid turning a rising chart into a claim of realized profit.
Our first question: what counts as learning?
A trading agent can change its rules every day and still fail to learn. It may react to the last loss, imitate a lucky paper position, or choose a rule because it looked good on the same data used to design it. A more demanding question is whether a proposed change improves later, independent decisions after the real cost of entering and exiting.
Researchers have shown why this matters. Bailey and coauthors describe a framework for estimating backtest overfitting risk: selecting the most attractive result among many tests can give a misleading impression of skill. A recent walk-forward study evaluates hypotheses on successive unseen periods with transaction costs and position constraints. Its authors report modest, statistically insignificant aggregate results. The study illustrates the value of reporting an uncertain outcome plainly.
How we would test a change
Start with one frozen rule and a comparison rule. Record what information existed when each signal appeared, the intended size, the quoted entry and exit routes, estimated costs, and the choice to trade or reject. Both rules should see the same incoming opportunities. If a new filter avoids one loss, its record must also include promising trades it skipped. Count independent closed trades, not repeated price marks on the same position.
For each version, compare realized net profit and loss, independent closes, average gain and loss, drawdown, and how often executable cost evidence is complete. Put paper outcomes beside paper outcomes and funded fills beside funded fills. For early-moving tokens, an apparent price rise means little if the intended position cannot be sold at the necessary size. Missing sell evidence is itself a research finding.
What comes next
These are proposed evaluation rules, not a claim that our current agents have achieved profitable learning. Our next experiment is to freeze a baseline and one challenger, expose them to the same candidate stream, and audit every independently closed outcome after costs. We will publish what improved, what failed, and what remains unknown. That is the standard this series will try to keep.