TechnologyMedium

3 min read · 514 words

What the seat actually does

Every seat on this map depends on software somebody wrote: the pricing library a structurer calls, the risk engine that runs overnight, the ledger operations reconciles, the app a retail customer opens. This seat builds and runs it.

The constraint that makes it different is that the system is live. A market does not pause for a deployment, a settlement deadline does not move, and a mispriced instrument is a real loss rather than a bug report. That is why finance technology is unusually conservative about release and unusually intolerant of silent failure — the two go together.

• Pricing and risk libraries — the code behind the numbers on every calculator a desk actually uses.
• Market data — sourcing it, cleaning it, and storing it so yesterday's number can be reproduced.
• Trading and booking systems, including the latency- sensitive end of them.
• Everything that is still running — the older the system, the more of the firm depends on it.

A day, and where it goes

• Overnight batches — what failed, what ran late, and whether the risk numbers were ready before the desks arrived.
• Releases, and the question of whether this one can be rolled back if it is wrong.
• Data. A missing or bad price feeds every downstream number, so it is found here or nowhere.
• Support — the seat next to a desk, which is where most requirements actually come from.

What it is measured on

• Availability during market hours, which is not the same as uptime averaged over a month.
• Correctness — whether the number the system produced was the number it should have produced.
• Time to deliver a change, against a business that measures opportunity in days.
• Incidents and their repeats. The same failure twice is the only one that counts as a process problem.

What it touches on this site

• The mathematics — valuation, volatility and curve construction, which are this seat's specifications.
• The seats it serves — quant, trading, risk and operations.
• Where the microseconds matter — market microstructure.
• When it goes wrong — 2010, where the interaction of several correct systems produced an incorrect market.

How it goes wrong

• Silent wrongness. A system that crashes is a good day; a system that returns a plausible wrong number is the expensive one.
• A release nobody can undo. Changes that write data are not symmetric with changes that read it.
• A spreadsheet as production. The tool built to bridge one gap becomes the thing the desk cannot operate without.
• Data lineage nobody can reconstruct. If yesterday's number cannot be reproduced, no dispute about it can be settled.

Concepts to master

• Reproducibility is a requirement. A number that cannot be regenerated from stored inputs cannot be defended to anybody.
• Floating point is not arithmetic, and money is the domain where that stops being a curiosity.
• Latency is a distribution, and the tail is the part that loses money.
• Every model has a domain — see valuation for what the numbers assume.