I run eight trading accounts in parallel to measure a single idea

Say you add a clever feature to a trading strategy. A news reading model, a memory of past mistakes, a smarter exit rule. A month later the account is up two percent. Did the feature do that, or did the market just go up? You genuinely cannot tell. One account gives you one story, and the market wrote most of it.

This is the problem that shaped the core of Aegis. I could not A/B test the way you would test a landing page, because I only have one market and it is the loudest variable in the room. So instead of testing one account harder, I ran more accounts, and I made them differ by exactly one thing each.

What the cohort is

Aegis runs a cohort of eight paper accounts at the same broker, each seeded with the same small amount of money, all trading the same entry signal on the same candidates on the same days. The only things that change between them are the capabilities I am trying to measure, added one rung at a time.

There are two rows and four columns.

Read across a row. The baseline account trades the raw rules with no model at all. The next account adds the news reading veto from the first post in this series. The next adds a per candidate context step. The last adds a memory of confirmed lessons. Each step is one flag flipped on, and nothing else moves.

Read down a column and you get the second experiment for free. The top row exits with a fixed stop and a fixed target. The bottom row keeps the same stop but scales its profit target to how much that particular stock actually moves. Same entries, same everything else, one different exit. So the columns isolate the model, and the rows isolate the exit.

The rule that makes it work

The discipline is boring and total: change exactly one variable per arm. The bottom row must mirror the top row in every single respect except the exit, or a difference between them could be a dozen things instead of one. The memory account must be identical to the context account except that it is allowed to use lessons, or "did memory help" becomes unanswerable.

It is the kind of rule that is easy to write and hard to keep, because every time I wanted to improve the whole cohort I had to resist, or I would have contaminated the very comparison I was running. A change that helps all eight accounts also erases the reason I have eight.

What eight accounts have shown me

Two things, both useful, neither flattering.

The model accounts and the baseline account sit within a fraction of a percent of each other. Across the whole run, the news reading veto has almost never been the difference. That is a real answer to a real question. I was measuring whether the language model earns its place, and so far, in this role, it does not.

The volatility exit row has run slightly ahead of the fixed exit row. Not by much, but consistently, and consistency across the same entries is exactly what this design is built to detect. It is the one hint I have that a piece of the strategy is worth changing, and I only trust it because the two rows were identical apart from that piece.

What it costs, and what it cannot fix

Eight accounts is eight times the bookkeeping, eight sets of positions to reconcile against the broker every day, eight ways for a fill to go missing. A lot of the engineering in Aegis exists just to keep these accounts honest and comparable, which is a later post.

And the design has a ceiling it cannot raise on its own. If the underlying idea has no edge, running eight clean copies of it just tells you that, precisely and repeatedly. Which, to be fair, is the job. A clean experiment that returns a null result is not a failure of the experiment. It is the experiment working, and saving you from believing a story the market happened to write.