Privacy policy. Plain terms, no surprises.
What we collect, why, who processes it and how long it stays. Zero customer portfolio data is used for model training. Last updated November 2026.
Who we are
Nyquist is a quantitative-based financial operating platform, operated by its founding team. Company details, registered jurisdiction, and corporate disclosures are published on the legal and security registries.
What we collect
We collect minimal data required to provide institutional services: contact information for pilot applications, session tokens for security, and operational audit logs when running live books.
For signed-in accounts we also keep a record of how each sign-in was made (password, passkey, or your firm’s own identity provider) and of the administrative actions taken in your firm’s account, so that the firm’s administrators can review them. When you sign in through your firm’s identity provider, we receive from it your verified e-mail address and name; we do not receive your password there. When you add a passkey, we store its public key, a name you give the device and a usage counter; the private key never leaves your device.
When a firm administrator invites you, we keep your e-mail address, the invitation’s expiry and a one-way hash of the one-time code (the code itself is not stored). The invitation is deleted with your firm’s data if the firm leaves the platform.
When a firm administrator proves ownership of an e-mail domain, we store the domain name, a random verification token and the time and outcome of the last check. To check it, our server looks up the domain’s public DNS records; no personal data is sent in that lookup.
How we use it
Your data is strictly used to deliver quantitative risk calculations, portfolio attribution, and model execution. We never sell, monetize, or share your data with advertisers or third parties.
Analytics & session replay
Two tools measure this website. Cloudflare Web Analytics counts page views and load times in aggregate and sets no cookie. Umami runs on our own Azure server, so its data never leaves our infrastructure. PostHog is not loaded on this website.
The signed-in platform and the public demo load those two tools as well. They also use PostHog (PostHog Cloud, US region) for product analytics: which pages and features are opened and in what order, clicks on interface elements, JavaScript errors and page timings. In the demo this includes the guide: which steps you open or skip, the role you pick and the routes and terms you open. PostHog keeps a random identifier in your browser’s local storage so page views from one browser join into one visit; it does not follow you across other sites, and we set no advertising cookies. Analytics events are stored by PostHog in the United States, so this data leaves the EU. Signed-in users are identified to PostHog by an internal account number only, never by name or e-mail.
Sessions are recorded only on the public demo (demo.nyquist.pro). There PostHog records a replay of the pages you open and what you click, type and scroll, so we can see where the product is confusing or broken. The demo needs no account and is meant for public data: do not enter positions or figures you want to keep private. Password, login, e-mail, phone-number and one-time-code fields are masked and never recorded; other fields you type into are. Recordings are stored by PostHog in the United States.
The demo assistant (Nyx) sends the question you type, up to the last six turns of that chat and public SEC 13F filing data to a third-party language-model provider (NVIDIA, Cerebras or OpenRouter, see the sub-processors page) and shows you the answer. Nothing from the chat is saved on our servers. To enforce a daily limit we keep only a per-day count of questions under a one-way hash of your IP address, deleted after two days. Do not type confidential information: the session replay above also covers the assistant.
The assistant on the signed-in platform records each answer of the language model, on our own server, as an AI verdict in the ontology so that it can be checked later in Evidence & Replay: the text of the answer, the model that wrote it, the run records it was given as sources, and for each figure in the answer whether it comes from one of those runs, from your own question or from neither. The question you typed and the earlier turns are not part of that record, nor of the run record written beside it (language, which sources, outcome of the check); they stay only in the chat session, which expires after 48 hours. Only you (the person who asked) and platform administrators can open the record in Evidence & Replay: a colleague in the same organisation does not see your answer in the list, by its identifier or in a search. Nothing is recorded for the public demo, for a question the assistant declined because it asked for a trade, when your organisation’s AI policy does not allow the call, or when the model fails.
The signed-in platform at nyquist.pro does not record your session: PostHog receives analytics events there, not a replay of your screen.
Error and speed reports
When a page of the platform or of the public demo breaks in your browser, the browser sends a short error report to our own server, not to a third party: the error text and its technical trace, the page path, the version of the application, the code of the page, and your browser’s user-agent string. Reports of content blocked by the page’s security policy are sent the same way. The report is sent without your sign-in, so it is not linked to your account.
Before a report is stored, the server removes anything that looks like an access token, key, password or e-mail address. We do not store your IP address with the report. Reports are kept for 30 days and then deleted automatically. We use them only to find and fix defects.
Once per visit to the platform or the demo, the browser also sends our own server three page-speed measurements: how long the main content took to appear, how much the layout shifted while loading, and how quickly the page responded to clicks and typing. With them go the page path, the version of the application and whether the device is a phone. This report carries no account, no cookie and no user-agent string, and we do not store your IP address with it. Speed reports are kept for 30 days and used only to find slow pages.
Shared links
When you copy a link to a page of the platform, the values you entered on that page and the instrument or portfolio you selected are encoded in the link. Anyone who has the link can read those values, so share it as you would share the calculation itself.
If the values are too long for the address bar, we keep them on our own server for 180 days after the link was last opened, so that the short link can restore them for the colleague you send it to. Only signed-in users of your own organisation can open a stored link. A stored link keeps the ID of the account that created it, and nothing else about its author. We do not put passwords, API keys, tokens or e-mail addresses into a link: fields that look like them are left out, and the page tells you how many.
Saved workspaces
When you save a workspace, we keep on our own server the pages you had open (their addresses, which carry the instrument or portfolio you selected and the values you entered) and the arrangement of your panels, under your account and inside your own organisation, until you delete it. A workspace you share is opened to the colleagues of your organisation you pick by name from a list of its active members (the list shows only the display names that members have set, never e-mail addresses, and is never shown to anyone outside the organisation), for reading and for copying: a colleague can make a copy of their own, and your original is not changed. You can also open a workspace to a team you belong to (a named group of people of your organisation, see “Notes, mentions and teams”), for reading or for editing: every current member of the team gets that access, and loses it when they leave the team or you stop sharing. A team can never delete a workspace or share it further. If your account does not belong to an organisation, workspaces cannot be shared at all. You can stop sharing or delete the workspace at any time. Each share, copy and deletion is written to the organisation’s audit log. We do not store passwords, API keys, tokens or e-mail addresses in a workspace: a workspace that contains them is refused.
Notes, mentions and teams
When you write a note on an instrument or another object, we keep on our own server its text, the object it is attached to, the ID of the account that wrote it, the time, and the IDs of the colleagues you mention, until you delete it. Editing a note adds a new version; the earlier versions are visible to you only. You choose who can read each note: only you, a team you belong to, or your whole organisation. Nobody outside your organisation can read it, whatever you choose. If your account does not belong to an organisation, notes stay private and nobody can be mentioned. When you delete a note, the text of all its versions is erased. A note can link a number to one of your own calculation runs. For each link we keep with that version of the note the value our server read from your run when you saved it, its unit if the result names one, the time of the result, and the ID and method of the run; the value is not copied into the text. Whoever can read the note sees these recorded values, but not your run. When you check the numbers of a note, we read your runs again and show the recorded and the current value side by side without changing the note. Deleting the note erases these records too. A mention reaches the colleague as a notification in their notification center and in the channels they have switched on (Telegram, e-mail, Slack); that notification says who or what was mentioned on which object but never carries the text of your note. We do not send note text to analytics, and the note panel is excluded from automatic click capture and session recording. A team is a named group of active people of your organisation that your firm administrator sets up: we keep its name and the IDs of its members. Teams do not exist for accounts outside an organisation. You can open your watchlist to a team you belong to, for reading or for adding and removing symbols; the list then shows to its members in your order, and you can stop at any time. Deleting your account erases the notes you wrote and takes you out of your teams.
Investment committee
When you hold a committee meeting in the platform, we keep on our own server its title and date, the quorum rule, the people you invite with their roles (chair, member, secretary, observer), the agenda items with the pack, run or memo each points to (its title and a digest of its content at the time you added it), every vote with the ID of the account that cast it and the time, the outcome of each item, and the conditions the committee attached, each with its text, due date, the object it refers to and the colleague who owns it. Who may chair a meeting and who may vote is decided by two roles of your organisation (committee chair and committee member) that the administrator of your organisation grants in the firm console. A voter may add a short reasoning of up to 500 characters to a vote; it is shown to the people who can read the meeting and is printed in the minutes. When the chair is handed over during a meeting, we record who handed it to whom and when. Votes, outcomes, chair hand-overs and every hand-over or completion of a task are only ever added to the record, never changed or removed; a vote can be revised until the outcome of that item is recorded, and the earlier vote stays in the record. Only the people invited to a meeting and the administrator of your organisation can read it. A colleague who is given a task sees that task, not the meeting. If your account does not belong to an organisation, a meeting is private to you and nobody else can be invited. The minutes (a PDF) are put together from these records only; no text is written by a model, and each item carries a hash of its records. The chair of the meeting can sign the closed minutes: the signature record keeps a copy of the minutes as they were signed (the people and roles, the items with their outcomes and the conditions with their text) and the ID of the account that signed, and it turns stale if the minutes change afterwards. When a task is assigned or handed over, and when it is overdue, its owner gets a notification in their notification center and in the channels they have switched on (Telegram, e-mail, Slack); an overdue task is also reported to the chair of its meeting in the same way. The notification says that a task was assigned or is overdue and gives its due date, never the text of the condition or the title of the meeting. The meeting pages are excluded from automatic click capture and session recording, and nothing from a meeting is sent to analytics. Meetings are part of the archive of your organisation’s data and are erased with it on exit, except the votes, outcomes, chair hand-overs and task events, which stay as an append-only journal: they carry no text except the reasoning of a vote, which is erased on exit while the vote itself stays.
Cards sent to your chat
The administrator of your organisation can connect a channel of the firm’s own Microsoft Teams orSlack by giving us the address of an incoming webhook that the chat service issued. We keep that address on our own server, encrypted, until the administrator removes it; it is never shown again, not even to the administrator. A member of the organisation can then send a card about the page they are looking at to one of those channels. The card contains only what that person chose to send: the page title and code, a link to the view, the time and the run it refers to, up to eight figures they ticked, and a short note. We add nothing else, and a card that contains a token, a key or an e-mail address is refused. The card goes straight from our server to the Teams or Slack address of your organisation, with no other party in between; the service that receives it (Microsoft or Slack) then handles it under the firm’s own agreement with it. We keep, in the organisation’s audit log, who sent a card, to which channel, when, and a fingerprint of its content, but not the figures. The administrator can allow a channel to carry links only, can forbid figures in cards for the whole organisation, or can require presentation mode, which hides figures; in both cases a card carries the link alone, whatever a channel allows. If your organisation shares only signed-off figures, figures of a view without a valid sign-off are not sent. Cards do not exist for accounts outside an organisation. Deleting the channel stops all sending to it.
Saved parameter sets
When you save the values of a page’s Parameters tab as a named set, we keep the set on our own server until you delete it: its name, the values you entered, the page it belongs to and the ID of the account that saved it. By default only you can open it. If you share it, every signed-in user of your own organisation can read and apply it, and only you can change or delete it. If your account does not belong to an organisation, a set cannot be shared at all. A set does not carry the instrument or portfolio you selected. We do not put passwords, API keys, tokens or e-mail addresses into a set: fields that look like them are left out, the page tells you how many, and the server refuses a set that still contains one.
Page view of a run
When you start a calculation from a page while signed in, we keep next to the run record the values of that page’s Parameters tab and the object you selected, in the same form as a shared link, up to 64 KB. It lets you reopen the run on its page. Before the run is written, the view waits for up to 15 minutes in a temporary store on our own server that only your account can read; the run record is visible only to you and stays until you ask us to delete it. We do not keep passwords, API keys, tokens or e-mail addresses in a view: a view that contains them is refused and the calculation runs without it. Demo guests and API keys send no view.
Notebooks from a page
When you press Open in Studio on a page, we keep on our own server, in a temporary store that only your account can read, the view of that page (the values of its Parameters tab and the object you selected, in the same form as a shared link, up to 64 KB) and up to eight requests the page sent to us in its last calculation: the path, the query and the body, that is the numbers you entered, together with a fingerprint of the inputs of the page’s own run. Everything expires 30 days after you last used it; each use starts the 30 days again. Nobody else, in your organisation or outside it, can read it, and another account gets the same “not found” as for a record that never existed. We do not keep passwords, API keys, tokens or e-mail addresses in it: a request that contains them is refused and no notebook is made. The notebook is built from this record whenever you or your Studio session ask for it. Our server does not run it: the code runs in Studio under your own session key, and the calculation it asks for runs on our server like any other. When such a run is recorded, the run record notes the page it came from and whether its inputs equal those of the page’s own run; only if they are equal does the record carry the page view, so the run can be reopened on that page. The run record is visible only to you. Demo guests have no such notebooks.
Chart drawings
When you draw on a chart (a trendline, a channel, a horizontal level, a Fibonacci retracement, a rectangle or a note), we keep the drawing on our own server until you delete it: the symbol, the points you placed (time and price), the label or note text you typed, its tone, whether you locked it and the ID of the account that drew it. Only you can read, change or delete a drawing; it is not shared with your organisation, and it is not tied to a timeframe, so it is there on another screen or device. Lines, levels and projections are calculated by our server from your points; the chart only paints them. Drawings are part of the archive of your organisation’s data and are erased with the account on exit. We do not accept labels that look like passwords, tokens or e-mail addresses.
Holdings monitor
The holdings monitor builds a calendar for one of your books and can watch it. To do that our server sends the tickers (and, for bonds, the identifiers) of your positions to public and market-data sources: the SEC’s EDGAR service for filings, an earnings-calendar feed for the largest positions, and the exchange’s data service for the terms of bonds. Those sources see our server’s address and the ticker; they do not receive your quantities, your account or your name. News is matched on our server against a public news wire by ticker, issuer name or LEI; the wire itself is fetched without any data about you. When you create a watch rule we keep on our own server the book it belongs to, the rule’s parameters (the filing forms or the number of days and event types you chose), whether it is on, the time it was created and the ID of the account that made it. For each alert that a rule has sent we keep a marker (rule, filing number or event identifier, symbol, title and time) so that the same alert is not sent twice. Only you can read, change or delete your rules; the alert itself reaches you through the channels you switched on in your notification settings and appears in your notification centre. Rules and markers are part of the archive of your organisation’s data and are erased with the account on exit.
Morning book brief
The morning book brief compares one of your books with the previous brief. For each brief we keep on our own server a snapshot of the book (the net asset value, each position’s symbol, quantity and price, the result of the nightly stress battery if one exists for that book, and the limits that apply), the difference to the previous snapshot, the events of the day and new filings for your positions, the list of what we could not find out, a short text, a hash of the inputs, the outcome of sending it and the ID of the account that owns it. If you switch on the daily delivery we also keep the time zone and the local time you chose. Only you can read your briefs and settings. The difference is computed on our server by fixed rules. The short text is written by a third-party language-model provider (see the sub-processors page), which receives the figures of that one difference (net asset value and its change, symbols and price moves of the largest movers, symbols of positions added or closed, the symbols and form types of new filings, the count of events, scenario names and limit names) and not your name, e-mail, account or organisation. The text may only restate those figures: a text with a number that is not in them, or with advice, is discarded and the table is shown without it. Sending the brief goes through the channels you switched on in your notification settings, and a copy appears in your notification centre, as described under Alert delivery. Briefs are kept for 13 months unless the administrator of your organisation chose a longer period or to keep them indefinitely in the firm console; once a day our server deletes the briefs older than that period, and a deleted brief cannot be restored. Briefs and settings are part of the archive of your organisation’s data and are erased with the account on exit.
Nightly checks of your books
Once a night, after 01:30 UTC, our server runs two checks on the books of your organisation, on our own server and without sending anything about you or your books to a third party. First, the stress battery: four fixed scenarios applied to a book’s positions, weighted as they are held, with the risk model estimated from our own price history of those same symbols. For each book and day we keep the net asset value it was computed on, the tail loss of each scenario, a description of the model (the symbols, the dates of the price window and the number of observations) and a reference to the book. A book we cannot stress from its own data, because a position has no price, no risk class in our instrument reference, or too little history, is kept as a day with the reason in words and no figures. A bond position carries no duration of its own, so the server asks the public data service of the Moscow Exchange for the duration of that bond, naming only the bond’s public identifier; if the exchange does not publish one, the book is kept as a refusal with that reason. The same battery runs on the portfolios of the desk screen, and the tail loss of each scenario of a successful run is also written, without a user name, to the analytics store of our own server next to the other computed results of your organisation. The latest run of each book or portfolio is also held for seven days in short-term memory, where the morning book brief and the pilot checklist read it. Second, the UCITS 5/10/40 check described under Compliance alert queue, on books where the rule is switched on.
The history lets you compare today with the previous day; only the owner of the book can read it. It is part of the archive of your organisation’s data and is erased with the account on exit.
Tracked screens
When you switch on tracking for a screener page, we keep on our own server the list of identifiers (tickers, issuer or fund names, as they stand in the universe you entered) that passed the screen in your last run, the page it belongs to, the time of that run, your choice of notice and the ID of the account that set it up, until you stop tracking. Only you can read it. Each new run replaces the list and tells you which identifiers entered and which left; the platform does not re-run a screen by itself, so a change is seen, and a notice is sent, when you run the screen. If you switch notices on, the names that entered or left are written to your own notification feed and sent through the channels you connected, as described under Alert delivery. Adding the passing rows to your watchlist stores their symbols in your own watchlist, the same as adding them one by one. The universe itself is not kept by tracking. We do not accept identifiers that look like passwords, tokens or e-mail addresses.
Data connectors
An administrator of your organisation can connect the platform to a file the fund already receives: an HTTPS address, an S3-compatible bucket or an SFTP server that the administrator names. Our server then fetches that file on the schedule the administrator sets, reads it with the column mapping the fund confirmed, and books the rows into the fund’s own book. The file itself is read in memory and not kept. We keep on our own server: the connector’s description (host, port, path, bucket, schedule and, for SFTP, the fingerprint of the server’s host key), a log of each run (time, outcome, the file’s name, size and hash, counts of rows booked and rejected, never the values in the file) and the ID of the account that set it up, until the connector is removed; the run log is kept as an audit record.
The administrator can also run a test (a dry run). It fetches and reads the file exactly as a scheduled run does, but books nothing and changes none of the connector’s bookkeeping. It leaves one line in the run log, marked as a dry run, with counts only; the first rows of the file, shown to the administrator on screen, are not stored.
The login, password, key or token the administrator enters for the source is kept only encrypted, with a key held outside the database. It is never shown again, never written to a log and never returned by the platform; removing the connector deletes it. The remote host sees our server’s IP address and the credentials entered for it. The platform connects only to public internet addresses over HTTPS or SSH, checks the address after name resolution and again on every connection, and stops a fetch that is too large or too slow. For SFTP the administrator pins the fingerprint of the server’s host key when setting the connector up; a server presenting any other key is refused before the password or key is sent. What a connector brings in stays inside your organisation: it is not placed in shared market-data stores or caches and is not visible to other organisations.
Book reconciliation
When you store a statement from your fund administrator or custodian to compare it with your book, we keep on our own server, under your organisation, the statement as you entered it (positions with quantity, price, value, currency and the time of the price, its date, its cut-off moment and the name you gave the source), a hash of its content, and the ID of the account that stored it. A statement built from an import batch is read from rows already in your book. Each reconciliation run is recorded with the hashes of the statement and of the book positions it read, the tolerances used and the counts it produced; a record with the same hashes is also written to your analysis journal. Every break found is written to a journal with the values of both sides, the likely cause chosen by fixed rules, and every decision made on it with the ID of the account that made it and the note it wrote. A break is never deleted and the book is never changed by a reconciliation. If you switch on an alert threshold, a new break at or above it is sent to the channels you connected, as described under Alert delivery. Statements stay inside your organisation: they are not placed in shared market-data stores or caches and are not visible to other organisations. Statements are deleted when your organisation exits; the run, break and decision journals are append-only and stay, as described under Retention.
The break investigation copilot works on the same records. Its suggestions (the likely type, cause, matching pair and next action, each pointing at records on both sides) are computed on our own server by fixed rules, without a language model, and written to an append-only journal together with the ID of the account that accepted or rejected each one and the note it wrote. A tolerance rule you propose is kept with its definition and the ID of the account that proposed it, and acts only while an approver’s signature covers it. These journals stay in your organisation and, like the other journals, are not deleted when it exits. Only when you press Explain in words do we send the facts of that one break (the instrument, both sides’ figures and the name you gave the statement source, the trades it refers to, the likely cause) to a third-party language-model provider (see the sub-processors page) to word a short explanation. The explanation is checked against those figures and shown to you, and is not stored; our own daily limit on explanations is counted per account. If no model is available the suggestions still work and the page says so.
Operations panel & data incidents
The operations panel reads, when you open it, the records your organisation already has: reconciliation breaks, FIX session state, data connector runs, price challenges and open margin calls. It does not copy them anywhere; it shows them to the people of your organisation only. What it adds is kept on our own server, under your organisation, in two append-only journals: the timers your firm administrator sets for each type of task (hours), the assignments you make (the row, the colleague, the due moment, your note and the ID of the account that made it), and a mark that an overdue reminder was sent. An overdue reminder goes to the assignee’s own alert channels, as described under Alert delivery, and names the task type and the due moment, never the row.
When you mark a value or the results of a run as bad, we write a data incident: its name, the reason you gave, a reference to the value or run, and a snapshot of what we found that stood on it (references to dependent values, runs, notes, reports, book briefs and sign-offs of your organisation, found in the recorded provenance). Every later step — a new search, a replay of a run, a draft review, its approval by a person and the closing — is another event with the ID of the account that made it. Nothing is deleted or edited; these journals stay when your organisation exits, but their content is erased, as described under Retention. A replay is a new run made by the existing run replay under your own credentials; the old run stays. The draft review is computed on our own server by fixed rules. Only when you press Draft with model text do we send that incident’s facts (its name, your reason, the counts and the chronology) to a third-party language-model provider (see the sub-processors page) to word a short summary. The text is checked against those facts, shown next to them, and a person approves the review; our daily limit on model use is counted per account.
Margin calls & collateral
When you use the margin call and collateral operations page, we keep on our own server, under your organisation, what you enter and paste: the terms of each credit support agreement (thresholds, minimum transfer amounts, independent amounts, the rounding step for deliveries and the one for returns, and the schedule of eligible collateral with its haircuts for each side of the agreement), records of exposure, posted, received and in-transit collateral, free inventory and cash balances and flows (amounts, currencies, dates, asset identifiers and the source label you gave), and the text of each counterparty notice you paste or read from a PDF, with the fields read from it by fixed rules, the quote that supports each field, and the ID of the account that stored it. Every step taken on a call — the assessment against your records, the decision to agree, agree in part or dispute, the proposed collateral and the note that a transfer was made outside the platform — is written to a journal with the ID of the account that took it. When you take the terms of an agreement from a contract abstract you saved and checked (see Contract abstracts), the terms record also keeps, for each term, the reference of that abstract, the page, a short quote from the contract and the mark that a person verified it; unlike the abstract, which is stored encrypted, these quotes are kept with the terms as the other fields of the record are, and are erased with them when your organisation exits. The bank file prepared for a signed proposal carries settlement details only from a record of the settlement instructions registry (see Trade support) that an approver has signed and another person has confirmed through a second channel; the file is returned to you and is not stored, and the platform does not send it. No language model reads the notice; the platform does not send the grounds for a dispute or any payment instruction, a person does. When a notice arrives, and again when a confirmed settlement date is within two days, we send an alert to the account that holds the call, through the channels that account switched on (see Alert delivery); the alert names neither the counterparty nor an amount, and carries only the ID of the notice and the date. The result of each assessment — the amount called, our amount, the difference with its reasons and, once taken, the decision — is also written, as exact decimal strings, to the analytics store of our own server next to the other computed results of your organisation, together with a record of the run that names the account that made it. Records are never changed in place: a correction is a new row that points at the old one. These records stay inside your organisation: they are not placed in shared market-data stores or caches and are not visible to other organisations. They are append-only and stay when your organisation exits, as described under Retention, with the amounts, notice texts, names and identifiers erased.
A counterparty notice can also be read from a PDF. You pick the file in the Data tab of the page’s Parameters; it stays in your browser until you press Read PDF notice, and is then held in the memory of that one request only and is not stored, unless you turn on the switch described under PDF documents you choose to keep: not on disk, not in any store, not used to train anything. What we keep is what we keep for a pasted notice — the text read from the file, the fields and quotes read from it, a hash of the text and, for the file, only the numbers of the pages that could not be read; the file itself and its name are not kept. A page of the PDF that has no text layer (a scan) is sent, within the same monthly page allowance and with the same limits as described under Questions to filings and documents, to Microsoft Azure AI Document Intelligence in Sweden Central; pages that already contain text are read on our server and are not sent. If the allowance is spent or the counter is unavailable, no page is sent and the notice names the pages that were not read.
Trade support and settlement fails
When you store a counterparty’s trade confirmations to match them with your trades, we keep on our own server, under your organisation and the book you chose, each line as you entered it (instrument, side, quantity, price, currency, trade and settlement dates, counterparty, the counterparty’s own reference), the kind of source and the name you gave it, a hash of the content and the ID of the account that stored it. Matching is done on our own server by fixed rules, without a language model, and changes no trade. A settlement that did not take place is recorded as a journal of events: when it was opened, the owner label and due date you set, every letter drafted or marked as sent (the text of the draft, the channel you named), the instruction files exported, and how it ended, each with the ID of the account that did it. An event is never edited or deleted. Settlement instructions (account and bank details) you propose for a counterparty are kept in the same way under your organisation, with the source you named, the ID of the account that proposed them and of the account that confirmed them on a second channel, and the reference you typed; they take effect only while an approver’s signature covers them, and are not visible to other organisations.
The platform sends nothing to a counterparty or a settlement agent: it prepares a letter and an instruction file, and a person sends them. Only when you press Draft reminder letter do we send the facts of that one settlement (instrument, quantities and amounts taken from your records, dates, the counterparty’s name) to a third-party language-model provider (see the sub-processors page) to word the letter; it is told not to add any figure, the result is checked against those facts, and a figure not in the record makes us use our template instead. Bank and account details are never sent to the model. Our own daily limit on model use is counted per account. If no model is available the template is used and the page says so. Stored confirmation lines are deleted when your organisation exits; the journals of settlement events and of instructions are append-only and stay, with their free text erased, as described under Retention.
Corporate action notices
When you paste the text of a corporate action notice from your custodian (an MT564 message or a letter) or upload it as a PDF (picked in the Data tab of the page’s Parameters, where it stays in your browser until you press the button), the document is held in the memory of that request only and is not stored, unless you turn on the switch described under PDF documents you choose to keep. What we keep on our own server, under your account and the book you chose, is the fields read from it, each with a short quote from the document and its place (line or page), the format, the name you gave the source, a hash of the document, and the time. A field with no quote that stands in the document is kept as not recognised, with the reason. A letter in free text is read by a third-party language-model provider (see the sub-processors page): the text of the notice, nothing else, is sent to it, and its answer is accepted only when our own code finds each quote in the document and each value in its quote; an MT564 message is read by our own code without a model. A page of a PDF that has no text layer is sent, within the same monthly page allowance and with the same limits as described under Questions to filings and documents, to Microsoft Azure AI Document Intelligence in Sweden Central.
The quantity you held on the record date is computed on our own server from your book; nothing about the notice is sent to another organisation. An election you prepare is kept with the option, the quantity, the owner label and due date you typed and the ID of your account, and is submitted only through an approver’s signature; the platform does not send the instruction to your custodian, you do. Our server regularly looks for a response deadline that is near or missed for a book of yours that has no signed decision, and sends one alert for it, to the channels you connected, as described under Alert delivery. The notice fields are deleted when your organisation exits; the election journal is append-only and stays, with its free text erased, as described under Retention.
Compliance alert queue
The compliance alert queue lists the firings of rules on your organisation’s books. It takes them only from our own records: orders stopped by a pre-trade gate (read from the audit of those orders), the check of a book against the UCITS 5/10/40 rule (run when a book owner asks for it, and once a night on every book where the rule is switched on), and the limits our risk watchdog trips, which are written at the moment the alert is delivered. For each firing we keep, on our own server and under your organisation, the rule, the value and the limit where the source recorded them, references to the order, book and instrument, references to the data it fired on with their as-of times, and the ID of the account whose activity or book it concerns. The queue is shown to the people of your organisation only.
Every decision is another event in the same append-only journal: confirmed, resolved, marked false (with the reason code and the words you wrote), an exception asked for or decided by a second person, a rule suppression asked for or decided, a reopening — each with the ID of the account that made it. A data cause for a false positive also opens a data incident, described above. Nothing is deleted or edited; the journal stays when your organisation exits, but its content is erased, as described under Retention. Only when you press Explain on a group of similar firings do we send those firings’ records (rule, value, limit, references, as-of times; never the words people wrote) to a third-party language-model provider (see the sub-processors page) to word a short summary. The text is checked against those records and never decides anything; our daily limit on model use is counted per account. An export for a period is built on request and sent to you; we do not keep a copy of the file.
Regulatory reports
When you prepare a regulatory report (today: the AIF report under Annex IV of the AIFMD, Section 1 of the SEC/CFTC Form PF, derivative position and trade reports under EMIR Refit ISO 20022 auth.030, and transaction reports under MiFIR RTS 22 ISO 20022 auth.016), we keep on our own server, under your organisation, for each draft: the file itself and its hash, the table of fields with the value, the status and the source of each (your book and trade ledger at the reporting date and the closing prices used, with the date of each price, or what you entered), the checks of the file against the regulator’s schema and the published rules, and what you entered for it: the identifiers of the fund and of its manager, the classification of each security (type, region, market, ISIN), the shares of investors and the prices you gave; for Form PF also the name, title, e-mail address and telephone number of the person who signs the form, as you typed them; for EMIR Refit and MiFIR RTS 22 also the legal-entity identifiers (LEIs) of the reporting, executing, submitting and opposite-side entities, trading-venue MICs, contract and UTI terms, and algorithm or client execution markers, as you entered them. The ID of the account that prepared the draft is kept with it. A draft is never changed in place: a new calculation is a new draft, and an older one stays at the version of the schema and the rules it was made with. The platform does not send a report to any regulator or to anyone else and does not act as the reporting party: you download the file and a person files it. A written explanation of what changed since the previous report is made by the “written explanations” function governed by your firm’s AI policy (see Model providers); it receives only the names of the changed fields and the numbers of the difference, never the file, and any figure in the text that is not in the difference is discarded. These records stay inside your organisation: they are not placed in shared market-data stores or caches and are not visible to other organisations. They are append-only and stay when your organisation exits, as described under Retention, with the file, the fields, the checks and the inputs erased.
Questions to filings and documents
When you ask a question to a company’s annual or quarterly report, our server reads that report from SEC EDGAR (sec.gov) for that one question and does not keep it: neither the report, nor its text, nor the answer is stored. The ticker or CIK you typed is sent to the SEC to find the filing; your account details are not. The answer is a quote found in the report by plain text matching on our own server, shown with its page and section; no language model reads your question or the report.
When you upload a PDF to ask it a question, the file is picked in the Data tab of the page’s Parameters and stays in your browser until you press the button; it is then held in the memory of that request only and is discarded when the answer is returned, unless you turn on the switch described under PDF documents you choose to keep: it is not written to disk or to any store, and it is not used to train anything. A page of the file that has no text layer (a scan) is sent, at most six such pages and only while a monthly page allowance lasts, to Microsoft Azure AI Document Intelligence in Sweden Central to be read; pages that already contain text are read on our server and are not sent. The recognised layout of a scanned page is cached on our server for 24 hours under a hash of the file, so asking the same file again does not send its pages twice. If the allowance is spent or the counter is unavailable, no page is sent and the answer names the pages that could not be read. See the sub-processors page.
PDF documents you choose to keep
The PDF you give to a question, to a corporate action notice or to a margin call is held in the memory of that request and discarded afterwards. Nothing is kept unless you choose to. The PDF input in the Data tab of the page’s Parameters has a switch, Keep this document. If you turn it on and pick a period of 30, 90 or 365 days, we store, on our own server and under your organisation only, the text of the pages as it was read (including the text recognised from a scanned page) encrypted — a separate key for every record, wrapped by a master key held outside the database — together with the file name, the SHA-256 of the file, the number of pages, the numbers of the pages that could not be read, and the screen it was kept for. The file itself is never stored, and nothing kept is used to train anything.
A kept document is listed in the same tab, and you can read it again without uploading it. Reading it again sends nothing to Microsoft Azure AI Document Intelligence and uses none of the monthly page allowance, because the text is already read; a page that could not be read when you kept the document stays unread, and to read it you upload the file again. The same file kept twice for the same screen is stored once, with the period you first chose, and one person can keep at most 40 documents. When the period ends the text is deleted by an automatic job, and you can delete it earlier from the list. A kept document is visible to the person who kept it and to no one else. It is not included in the organisation’s data archive as readable text: the archive lists it by name, page count and retention date only. It is deleted when your organisation exits. Keeping and deleting a document are written to the audit journal with the screen and the period, never with the text. Backups made before it was deleted still hold it until they expire, after 60 days.
Contract abstracts
When you ask the platform to read a contract (an ISDA Schedule, a credit support annex, a side letter, a fund document), the PDF is held in the memory of that request only and is discarded when the answer is returned. A page of the file that has no text layer is read by Microsoft Azure AI Document Intelligence on the same terms as described under questions to documents. The fields are found by plain text matching on our own server; no language model reads the contract. Each field is shown with its page and a quote from the contract, or as “not found”; a number is never filled in from anywhere but the text.
Nothing is kept unless you choose to. If you save an abstract, you confirm it in the dialog, pick a retention period of 30, 90 or 365 days, and we store, on our own server and under your organisation only, the fields and the quotes from the contract encrypted (a separate key for every record, wrapped by a master key held outside the database), the name you gave the agreement, the file name, the SHA-256 of the file and the marks you put on fields (verified or rejected, with the time). The file itself is never stored. When the period ends the abstract and its marks are deleted by an automatic job; you can delete them earlier. A saved abstract is visible to the person who saved it and to no one else, and it is not included in the organisation’s data archive as readable text: the archive lists it by name, version and retention date only.
Counterparty files
A counterparty file describes a legal entity, not a person: its legal name, LEI, jurisdiction, the risk tier and review interval you set, and an amount you propose as a limit. We do not ask for or store the names or identity numbers of individuals (owners, directors, signatories) as fields of the file. A document you add (a certificate of incorporation, a register extract, a licence) is read on our server like a contract: we keep its type, its issue and expiry dates, the page count and the SHA-256 of the file, and the quotes that show where each date was read, encrypted. The document itself is not stored. A quote may contain a name if the page does; for that reason quotes are never shown outside your own account.
Checks are recorded with their date: a sanctions search by the entity’s name, which runs on our server against copies of the public lists we have a live source for (the OFAC SDN list of the US Treasury, the EU consolidated financial sanctions list of the European Commission and the UK sanctions list of the UK Foreign, Commonwealth & Development Office). Our server downloads each whole file itself, from treasury.gov, webgate.ec.europa.eu and sanctionslist.fcdo.gov.uk: those requests carry no name and no data of yours. The UN Security Council consolidated list is switched off until its terms of use have been cleared: while it is off, our server does not contact the UN site (scsanctions.un.org) and a check records the list as “not screened: licence”, never as clear. Each check records, per list, the publication date the source gives and the SHA-256 of the file we compared against; a list that could not be downloaded or read is recorded as not screened, with the reason, never as clear. Other lists are shown as not connected. The check also includes a lookup of the LEI and its parents in the public GLEIF registry, and your own note of the publications you reviewed. The only thing sent outside is the LEI you entered, to GLEIF; the name you search never leaves our server, and your account details are not sent anywhere. A possible match in a sanctions list is closed only by a person, with a written basis; the platform and its models never close it. Checks and decisions are append-only journals: they stay with a pseudonymous account number if the organisation exits, and the basis text is erased. The rest of a file is kept until you delete it or the organisation exits. About 30 days before a document expires we send you one reminder in the notification centre (and on the channels you switched on), without the name of the counterparty.
Alert delivery
When an alert you set (a price level, a risk limit) fires, we send its text — what happened, the source and the severity — to the channels you switched on in your profile and only to them: your account e-mail address (through Resend), the Telegram chat linked to your account, and the Slack workspace you linked. The same record appears in your notification centre in the platform. It is stored on our own server, visible only to your account, for 30 days, together with what you did with it (marked read, acknowledged) and which channel accepted it or refused it. A critical alert you have not acknowledged is repeated up to twice, fifteen minutes apart. The quiet hours you set in your profile are kept with your preferences and hold messages to those channels; the record in the platform stays.
The platform also applies three built-in risk limits (portfolio drop, single-asset drop, volatility spike) to every book of yours. For a book you can set your own threshold and pause for a limit, or switch the limit off. We keep that setting — the limit, the book, your values and the switch — on our own server under your account until you reset it, and every change is written to your audit trail with the old and new values.
Scheduled reports
A report you build is kept on our own server under your account: its blocks, its schedule, the colleagues who receive it and the channels used (e-mail, Telegram, Slack). Recipients can only be members of your own organisation.
When a report runs, we render a PDF from numbers already recorded in your own runs and send it. The PDF travels as an attachment by e-mail through Resend; Telegram and Slack messages carry the report title, its time and a link, never its figures. Every run is written to a delivery journal: who was addressed, over which channel, whether it arrived or why not, and a copy of the PDF so you can download it again. The journal and the copies stay until you ask us to delete them.
Webhooks
An administrator of your organisation can name an HTTPS address and the events it should receive (an alert delivered to a colleague, an approval requested or decided, a book import, a reconciliation run that opened breaks, a run recorded for your organisation, a morning book brief built). When one happens, our server sends a signed request to that address. The request carries the event’s ID, its type and time, the ID of the object it refers to and a link to the API path where your own key fetches that object —never the figures or text of the event itself. For the last two events it adds a short label: for a run, its route (for example POST /api/risk/var), the moment it was recorded and whether it ended in success or error; for a brief, its day, whether it came from the schedule or was started by hand, and the ID of its run record. Nothing from a run’s inputs, its results, the text of an error or the text of a brief is sent. A run recorded by any person in your organisation is announced to the organisation’s address, so the administrator who names it sees that the run happened, not what was calculated. The receiving host sees our server’s IP address and that request. The platform connects only to public internet addresses over HTTPS, checks the address after name resolution and on every attempt, and follows no redirects. A failed delivery is retried up to six times with growing pauses and then abandoned.
We keep on our own server, under your organisation: the address, the chosen events, the note the administrator wrote, the ID of the account that set it up, one queue row per delivery (event ID, type, object ID and link, the short label above where the event has one, its state) and a log of every attempt (time, the status code your server answered with, a class of outcome and the duration — never the body of your server’s answer). The signing secret is shown once and kept only encrypted, with a key held outside the database. The subscription and the queue are deleted when the administrator removes the subscription or your organisation exits; the attempt log is append-only and stays, as described under Retention. Every change an administrator makes is written to your audit trail.
Support requests and feature board
When you write to us from the product, the ticket is kept on our own server under your account: the subject and the text you wrote, every later message (yours and ours), its status, and the attachments you chose. The form offers three, each switched on only by you: the code of the page you were on, the build version of the portal, and the last error your browser reported (cleaned of tokens, keys and e-mail addresses before it is sent). Nothing else is attached — no page parameters, no IP address, no browser details. Our answer reaches you in the notification centre and on the ticket itself; it is not e-mailed. Tickets stay until you ask us to delete them or your organisation exits.
The feature board lists requests the team has put there. When you vote, we keep one row with your account and the request; the number you see is a count of those rows, and nobody sees who voted. Removing your vote deletes the row. Votes are deleted when your organisation exits.
BI feed and Office live links
A key you mint with the BI feed scope can only read a read-only feed of your own runs, their result values, your book positions and your reconciliations, and nothing else on the platform. You give that key to the BI tool you connect (Power BI, Tableau); from then on that tool receives the rows under your responsibility, and the figures leave our servers for the tool’s host. Every row says which run it comes from, as of when and from which source, and a figure that is old or unpriced is flagged, never replaced. We keep no copy of the rows sent: the only trace on our side is the key itself (stored as a hash with its scope label and the time it was last used) and the ordinary request log. The administrator of your organisation can switch the BI feed and the Office live links off for the whole organisation: a switched-off feed refuses every request, with the reason, and the keys already issued are kept but read nothing until the administrator switches it on again.
The Word and PowerPoint add-in sends us a run ID or a link to a leaf view and your API key, and nothing from your document. The table you insert lives in your document, with the run ID, time and source written under it. “Refresh” asks our server to run the calculation again as you; the new run is recorded in your own run journal like any other run, and your document changes only when you accept the new values. The add-in keeps your API key in the local storage of its pane on your device.
Research survey
The survey at discovery.nyquist.pro asks about your institution type, your role and, optionally, its size, then about the tools you run, what they cost and what you would prioritise. Answers are stored as the codes of the options you chose, plus any free text you type, in our database on Microsoft Azure in the EU. With each response we also store the survey language, your browser language, the address of the page without parameters, and the tags of the link you came from: a code identifying the invitation and, for our e-mail invitations, the institution it was addressed to. The tags show us which invitations were answered; they never appear in the published report.
We do not store your IP address with a response. The survey’s web server keeps an access log (address, page requested, browser name) in rotating files of up to 10 MB each, five files at most, and the rate limiter counts requests per address for one hour. The page loads Cloudflare Web Analytics (page loads, no cookie) and Umami on our own server, which receives the steps you open and the options you choose together with the link’s tags, never free text or contact details. When a response arrives, our team is notified in a private Slack channel and by e-mail (through Resend) with the institution type, role, size range and the free-text answer about the most painful workflow; contact details are not part of that notice.
After you submit, you may leave an e-mail address, and optionally a name, company and a LinkedIn profile or phone, and tick what it may be used for: the report, a conversation about Nyquist, or both. These go into a separate table, are used only for the purposes you ticked and are kept under the retention stated in the consent text shown with the form. Results are published in aggregate only; segments with fewer than five responses are merged or dropped. There is no automatic deletion date for answers yet. To remove a response and the contact details tied to it, write to contact@nyquist.pro from the address you left; an answer sent without an address cannot be traced back to you and cannot be removed one by one.
Survey e-mail
We invite people who work at financial institutions to take a short research survey. We write to business addresses that the institution itself publishes on its own website. The e-mails are sent through our e-mail provider, Resend.
Where the provider reports them, we record what happened to each message: delivered, opened, link followed, bounced or marked as spam. For each event we store the recipient’s address, the type of event, the time, which link was followed and the message details the provider reports (such as subject and sender). We do not store the IP address or the device details of the person who opened or clicked: they are removed before the record is written. The provider keeps its own records under its own privacy policy. We use these records to see how the survey invitation performed. To be taken off the list, reply to the e-mail or write to the address in Contact; we will remove your address and delete the records tied to it.
Model providers & zero-training guarantee
Strict zero-training policy: Customer financial positions, portfolio weights, trades, and proprietary theses are never used to train or fine-tune models. In dedicated VPC and on-premise deployments, model inference is configured by contract.
AI policy of your firm. An administrator of your firm decides what the AI functions may do: switch AI off, allow only some model providers, switch the assistant, natural-language queries and written explanations on or off one by one, and say whether data from your firm’s books (positions, weights, the results of your calculations) may be sent to a hosted model at all. Until an administrator allows it, that data is not sent, and a switched-off function answers “disabled by your firm’s policy” with the reason, not with an error. Every change of the policy is written to your firm’s activity journal. Requests we send to OpenRouter on behalf of a firm carry the setting that forbids the provider to store or train on them; for NVIDIA and Cerebras no such per-request setting exists and the terms of our agreement with the provider apply.
Question to calculation. When you type a question in the command bar while signed in, fixed rules on our own server prepare a plan: the sheet, and the numbers read from your question word for word. The question and the plan are kept for 24 hours so that a run you start from the plan can cite them; they are written to the run record only if you start the run. Only when the rules find no calculation, and your firm’s AI policy and the sign-in allow it, the text of the question (nothing else) is sent to the same model providers as above to propose a plan. This is a function of its own in the firm’s AI policy (“Calculation plan”): an administrator can switch it off without touching the others, and the rules keep answering. The model names a calculation and the words of your question that hold each number; the platform reads the numbers itself and drops any that are not in your question. A plan proposed by the model opens nothing until you confirm it. The public demo never sends a question to a model for this. Such a call is recorded in the AI call journal below like any other.
AI call journal. For each call to a model made while you are signed in we record, per call: the ID of the account, the function, the provider and model, the number of tokens, the kinds of data the request carried (for example “the text the person typed” or “results of the person’s own runs”), the outcome, and whether the answer passed the check of its figures against the sources. The text of the question and of the answer is not recorded. A call that the policy refused is recorded too, with the reason. The journal is visible and exportable only to the administrators of your firm (an individual account has none, so its entries are kept but shown to no one), is append-only, and is deleted automatically after the retention period your firm sets (at least six months; one year unless set otherwise). It stays in your organisation when it exits until that period ends and, after the account is anonymised, carries only a pseudonymous account ID. Calls made without an account (the public demo) and by our own background jobs are not recorded in it.
Sub-processors
The third parties that receive personal or customer data — hosting, e-mail, analytics, model providers and the delivery channels you connect — are listed on the sub-processors page, with what each one receives and where.
Retention
Operational data is retained only for the duration of your active pilot or contract period, plus any retention period required by your regulator. Nightly database backups are kept for up to 60 days (3 days in hot blob storage and 60 days in archive storage). A contract abstract you chose to save is kept only for the period you picked (30, 90 or 365 days) and is then deleted; backups made before it was deleted still hold it until they expire, after 60 days. The same holds for the text of a PDF document you chose to keep.
When the contract of an organisation expires, every member sees a notice for 14 days and can work as before. After that the workspace is closed for saving: reading, calculations that store nothing (a VaR, a stress run, the price of an option) and exporting your own data stay available, while saving trades, books, settings, documents and signatures is refused, and nothing is deleted. Data is deleted only when the organisation asks to exit.
Your rights
Under GDPR, CCPA, and global privacy standards, you have the right to inspect, export, or delete your data. What the platform does today:
Your own account. In Profile → Permissions you can download your account data (profile, preferences and linked identifiers) as JSON, and delete your account. Deletion anonymises the account in place; it is not a hard delete, because the audit chain refers to it.
Your organisation’s data. The owner of an organisation can download everything the organisation holds from Profile → Data and exit: books, trades, runs and their inputs, rules, workspaces, import history and journals. It is one archive with a file in JSON Lines for each table and a manifest that gives the SHA-256 of every file, so you can check it is complete. Credentials and session tokens are never exported.
Leaving. The owner can request exit and confirm it with the organisation’s identifier (its slug) and password. After a waiting period of seven days the owner executes it. Then the organisation’s runs, rules, workspaces and other working data are erased and its users are anonymised. Committed trades, orders, execution events and audit journals are not erased: they are append-only, and the audit chain would break. They stay linked to a pseudonymous account number. Erased rows leave our backups when the last backup holding them expires, after 60 days. Files are deleted too: the reports, datasets and notebooks your people stored with us are removed from our file storage, and so are the register files you exported from the registers (Parquet files on our server, each kept in a folder of its organisation and its owner). The act gives the number of files and bytes removed, per storage area, and any error. Register files exported before the folders existed carry no owner, so such a file is deleted only if no other person’s record uses the same file name; the act counts the ones left, and you can ask us at the address below to remove them. A refusal of the storage is never reported as done, and the exit is not completed until the step is repeated and succeeds. The one exception is the PDF of each report mailing: it sits in the delivery journal and is kept indefinitely, linked to a pseudonymous account number. The profiles and events of the people who leave are deleted from PostHog as a separate step, by a request to its API. This step does not hold the exit: if PostHog refuses the request, or the platform has no analytics-processor key configured, the data in our database is still erased and the act says “not performed” for this step, giving the response code and no response body. The act has a button that repeats only this step; the repeat updates the same act. It works from the act shown when the exit is executed. If you have closed it, or the step keeps failing, send the request to the address below. We store an act that lists what was erased and what was kept, and the owner can download it.
Leaving the shared legacy organisation. People whose accounts belong to the shared organisation created before organisations existed cannot take it with them, because it holds other people’s data. Each of them can delete their own account from Profile → Data and exit, after the same seven-day waiting period and with the same act. Before deleting, they can download their own data from the same screen, at any time while the account exists, including during the waiting period: it is the same archive format as the organisation’s (a file in JSON Lines for each table and a manifest with the SHA-256 of every file), but it holds only the rows the person owns. Rows that belong to the whole shared organisation, and other people’s rows, are not in it, and credentials and session tokens are never exported. The archive is built when you ask for it and handed to you in the response; we keep no copy and no download link, and each download is written to the action log. Only the rows they own are erased, together with their files and their PostHog profile; their sessions and API keys stop working. Other people’s rows are not touched. Trades, orders, execution events and journals stay linked to a pseudonymous account number, as above.
Security
All data in transit is encrypted via TLS 1.3. Data at rest is encrypted via AES-256. Private deployments run inside on-premise or single-tenant VPC environments.
Contact
For privacy and data sovereignty inquiries: privacy@nyquist.pro
