Data, connectors and preparation
Data that says what it is, not what would look good.
A universe can come from a CSV you already have, a bundled index catalogue, or one of several providers — each carries a status, not a promise. Preparation runs in a fixed order with a policy at every stage and an audit of what it actually did. See /security for how credentials and network access are handled underneath this.
- Data providers
- 11From synthetic reference data to enterprise adapters, each with a stated status
- Bundled index catalogues
- 15Coverage marked full, partial or composite — never complete by default
- Preparation stages
- 7Fixed order: duplicates, price basis, missing values, calendar, outliers, resampling, minimum history
01What you actually do here
Five things, in the order the page presents them.
- 01
Pick a universe
Choose an index or catalogue, a provider and a date range, then tick the securities you actually want. A new project starts with no universe — nothing is seeded from a preset's headline constituents until you apply it.
- 02
Or bring a CSV
Upload requires date, ticker, open, high, low, close and volume; adj_close is optional and defaults to close. Rows with an unparseable date, ticker or close are dropped and counted, not silently discarded. The upload reconciles your project's declared universe and benchmark against what the file actually contains, and warns if the benchmark is missing or average observations per security fall under 260.
- 03
Prepare the data in a fixed order
Duplicates, then price basis, then missing values, then calendar alignment, then outlier treatment, then resampling, then minimum history — the order is not user-configurable, because a later stage assumes an earlier one already ran. Each stage carries its own policy and its own audit line.
- 04
Build features on the prepared series
Every feature value at time t uses only data known at or before t. Features are drawn from a versioned registry, and a baseline run without engineered features sits alongside the engineered one so the comparison is visible rather than assumed.
- 05
Check what is actually connected
Settings reports each provider's status and where its configuration came from, without ever returning a credential value. Market data, macro-economic data and enterprise adapters such as LSEG and Bloomberg each show what they need before a run can use them, so nothing downstream rests on a connection that was assumed.
02How it works
The method, stated for the specialist.
A provider is data, or it is stated as unavailable — the resolver never quietly substitutes one for the other. The same discipline extends to the network access underneath it and to how a prepared series is allowed to move between stages. See /security for the session, credential and tenant-isolation model this sits inside.
- Eleven providers, honestly labelled
- Quant Compose Reference Data is deterministic, seeded and synthetic — never a basis for a claim about a real security. This project's dataset and Workspace CSV files serve whatever is already loaded. Yahoo Finance and Ken French Data Library need only internet access. FRED and Alpha Vantage are conditional on a configured API key, Alpha Vantage further rate-limited to 5 calls a minute. Polygon is conditional and does not adjust dividends on its daily aggregates. A custom REST endpoint is configured, not coded, requiring a registered endpoint. LSEG and Bloomberg are enterprise adapters that depend on licensed access the workspace does not supply; until that access is in place the data path says so rather than guessing.
- resolve() never falls back to synthetic
- A real provider that fails returns a failure, not a quietly substituted synthetic series. A fallback to synthetic data is worse than no data, because no data stops you and plausible data does not.
- An egress allowlist, not an open network
- Only approved hosts are reachable, HTTPS is enforced unless a host is explicitly opted out, and every redirect hop is re-checked against the allowlist. Private, loopback and link-local addresses are refused for every resolved IP — an explicit defence against a request being redirected somewhere it should not go. Requests carry a bounded timeout, a maximum response size and a maximum redirect count, and credentials are redacted before anything is logged.
- Fifteen bundled index catalogues, coverage stated
- From the S&P 500 and FTSE 100 down to a sample-only Russell 2000, each catalogue's coverage is marked full, partial or composite against the real index membership — never claimed complete when it is not.
- The preparation order is fixed for a reason
- Duplicates are resolved before a price basis is chosen, because a duplicate row would otherwise corrupt whichever basis you pick. Missing values are handled before calendar alignment, calendar alignment before outlier treatment, and so on through resampling and a minimum-history check — each stage assumes the one before it already ran, so the order is not a setting.
- Policies at each stage, and a tag on the result
- Missing values: drop, forward-fill, backward-fill or interpolate. Outliers: none, winsorise at 1%/99% on daily returns, or clip at ±4 standard deviations with the price rebuilt from the clipped return series. Calendar alignment keeps only dates every security actually trades on, and warns above 25% row removal. Every stage is tagged SAFE, REQUIRES CAUTION or MODEL-DEPENDENT so a policy choice reads as a decision, not a default nobody looked at.
- FX is a presentation concern
- Converting to a reporting currency uses a single spot rate per pair, and that conversion never touches the return series feeding risk, features, models, portfolios or backtests. Converting with a time-varying rate would be a modelling decision, not a display preference, and this workspace never makes that substitution implicitly.
03What you get out
Outputs, and what they let you defend.
Preparation and provider status are both recorded, so a downstream figure can be traced back to the data it actually rested on.
- A reconciled universe and benchmark, with the exact rows the upload rejected and why
- A stage-by-stage preparation audit: policy applied, rows affected, SAFE/CAUTION/MODEL-DEPENDENT tag
- Provider configuration status and source, visible without ever exposing a credential
- A point-in-time feature set with a versioned registry entry and a baseline comparison
- Watchlists and screens built from Market Explorer, ready to become a project universe
04What this does not do
Stated plainly, because it matters as much as what it does.
CSV only
Upload accepts CSV. There is no XLSX or JSON upload path.
LSEG and Bloomberg need your own licence
Both are enterprise adapters that depend on licensed access the workspace does not supply. Without it the workspace states that the source is unavailable rather than substituting another.
Synthetic data proves nothing about a real security
Quant Compose Reference Data is deterministic and seeded for exercising the workspace without a network call. It is never a basis for a claim about a real market.
Corporate-action check is detection only
It flags a price jump consistent with an unadjusted split or capital event. It does not correct the series for you.
Keep going
Related capabilities
Bring your own data, or start from a bundled index.
Beta researchers upload a CSV or apply a bundled catalogue, then watch the preparation audit before anything downstream runs.