Run it again and get the same answer
By Robert Yenokyan,
A committee cannot approve a number that changes when you run it again.
Getting the same number twice has to be built in from the start. You find out whether it was on the day somebody asks for that run again.
If you sign off what gets traded with client money, that is your problem long before it is the researcher's.
A research result is a claim about what would have happened. Somebody has to take it apart months later, usually somebody who was not in the room when it was made.
That only works if the run can be produced again exactly. Same inputs, same output, character for character, on a different day. Anything that comes back different has a cause, and somebody can point at it.
It gets treated as engineering tidiness. It is the thing that has to be true before anyone can sign.
So every research and simulation run we do leaves a record. Every fill, every order the venue refused and every cancel, time stamped in the order they happened. A model risk reviewer, an outside auditor and your own internal file read the same record instead of three reconstructions of it.
It also ends arguments. When two runs disagree, the record shows where they split.
So ask for this. Take a result from last quarter and have it produced again today, while you watch. Then ask what a difference would have meant.
A team that cannot do that is not being sloppy. It is asking you to take the number on trust, and that is the one thing your own reviewers cannot do.
We build and run the infrastructure underneath systematic trading, and the record comes with the work rather than as a favour on top. If a result at your firm cannot be produced twice, write to robert_yenokyan@bonton.ai.
Byte-identical is the precise version of "the same answer", and it is the useful one, because it means a difference of one character is enough to start the question. The record is of research and simulation runs, which is the layer where a result gets decided.