For a prop desk's risk officer
The behaviour behind a breach, read from the tradebook
Riskora reads a trader's own exported tradebook and describes how they behaved around their own wins and losses: how fast they re-entered after a loss, whether size changed after losing, whether losers were held longer than winners. It shows the arithmetic and names the rows in the trader's own file.
It is not a signal service, it does not connect to a broker, and it does not sit in the order path.
Whether any of these patterns predicts anything is unknown. This page proposes the test rather than asserting the answer, and states the decision rule before the data exists.
Why a prop firm, and not an academy
A prop firm has the one thing no amount of engineering can generate: a drawdown-breach outcome. Accounts pass or fail against a stated rule, the firm records which, and the record is unambiguous.
An academy can tell us whether a behavioural pattern recurs for the same person. Only a firm with breach outcomes can eventually tell us whether it corresponds to anything that costs money. Those are two different tests, and the second needs the first to have passed.
The honest version is unflattering. The breach-correlation question is the one a risk officer cares about, and it is the one we are furthest from being able to answer. The recurrence test comes first because it is cheap, it is already implemented, and a signal that does not even recur for the same person cannot correlate with anything.
What is measured, precisely
Six behavioural signals, each defined in the specification with its thresholds. Every threshold is a documented judgement call, not a number fitted to data, and the specification says so on its own face. For a risk officer that is the single most important sentence here: nothing has been calibrated against outcomes, because there are no outcomes in the dataset.
What the reader gets for each finding:
- the numbers it came from, in units a trader recognises
- the rows in their own file, by line number, so the finding can be checked against the source
- what could not be measured, and why, rather than a zero. A figure that could not be computed is reported as unavailable with a reason, never rendered as zero, because a zero reads as a clean record
The test, registered before any data
Pre-registered in the specification, dated 2026-10-23, written before any production data was queried, and already implemented.
The claim
For each signal, among participants with at least two readings where the signal was measurable in both: the probability that it fires at the later reading, given that it fired at the first, exceeds the base rate across the same participants by at least 15 percentage points, with at least 20 such participants for that signal.
The comparison against the base rate is the claim. "80% of flagged traders were flagged again" means nothing if 80% of everyone is flagged. Reporting the first number alone is the easiest way to mislead a buyer, and the test is written so that cannot happen by accident.
| Outcome | What we may then say |
|---|---|
| lift 15 points or more, n of 20 or more | "this pattern recurred for this cohort", with the lift and n stated |
| lift under 15 points, n of 20 or more | it must not be sold as predictive, and the thresholds need recalibration first |
| n under 20 | "not yet established" |
A failed test is a result, not a reason to move the bound. The 15 points and the n of 20 are not to be edited after seeing the data. If they turn out to be the wrong test, that is a new pre-registration with a new date, and the old one stays on the record.
What is true today
| Paying customers | Zero |
| Pilots ever run | Zero |
| The cohort layer | Built and automatically tested. Never used by a real cohort |
| Predictive validity | Not established. No test has had data to run against |
| Correlation with drawdown breaches | Never tested. Breach outcomes are not in the data at all |
| Regulatory status | None. Not a regulated entity, holds no licence or certification, and claims none |
| Price | Not set. If you ask in a meeting you will be told that rather than given a number |
There is no pilot customer to reference and no case study. A firm reading this would be the first.
Two disclosures you would otherwise find yourself
The discipline score's audit trail
The ORA-1 discipline score penalises a trader when the server refuses a trade, and restores points for a trading day with no violation. Until 2026-11-08 the record of those violations was written by the trader's own browser, even though the refusal itself was always decided by the server. A trader who interfered with that one request was refused the trade and docked the penalty, then recovered the same points as somebody who broke no rules; and a trader whose own machine flagged something the server had allowed lost points for a violation nothing enforced.
Both are closed. The server now writes the record at the moment it refuses. Any score history covering a period before 2026-11-08 reflects a partly self-reported trail, and we will say so rather than present it as enforcement history. A pilot starting after that date is unaffected.
Tenant isolation
The organisation boundary is written and not yet applied. The migration that adds the tables and policies exists and has been reviewed; applying it to the database has not happened. Cohort reads currently go through a privileged server connection. A boundary in a reviewed file is not a boundary in a database, and you should ask us for the applied state before a pilot handles your traders' data rather than taking this page's word for it.
What Riskora is not
Riskora never routes an order, never holds client funds, and has no execution path to an exchange. It gives no capital, no payouts and no profit share. It is not a signal provider and not an advisory. Trading inside Riskora's own terminal is simulated with virtual capital. Any broker connection we would discuss is read only. The architecture is privacy minimised by design rather than by policy: the tradebook is parsed in the trader's browser and the raw rows are never transmitted.
What a pilot looks like
Your traders export the tradebooks they already have, as CSV or Excel. Each one runs their own file in their own browser. What reaches our server is the derived behavioural numbers and only those on a fixed list, never the trades. You get the desk view: which behaviours are common, which are concentrated, and what they cost, with small groups suppressed so no individual is identifiable from it. Then you decide. There is nothing to cancel.