Catalog
Datasets
Data quality is not a support function for a backtesting platform — it is most of the result. A flawless engine running against a survivorship-biased universe produces a confident, precise, wrong answer. This is the catalog we are assembling, and the state each piece is actually in.
Status: two live, three planned, none purchasable
Two datasets below are real and running: the macro series (published to /macro, regenerated by the engine) and the survivorship-free equity mirror, which is marked internal because it is licensed vendor data we are not permitted to redistribute. The remaining three are planned — sources are named so the intent is checkable, but availability, licensing, and access mechanics are unsettled.
Live does not mean purchasable. There is no account system, no API, and no delivery mechanism of any kind today, so no dataset here is available to anyone outside the operator regardless of its status below.
Survivorship-free US equities
Live (internal)The single most important dataset in the catalog. A point-in-time universe that retains companies which were later acquired, delisted, or went to zero is what separates an equity backtest that means something from one that has quietly pre-selected for survivors.
Includes the delisting and corporate-action history needed to reconstruct the tradable universe as it stood on any past date, rather than as it stands today.
Full-market daily bars
PlannedThe base layer everything else sits on. Breadth matters more than depth here — a strategy screened against a hand-picked subset of tickers is being tested on a universe that was already curated by someone with hindsight.
Split- and dividend-adjusted at the data layer, with the adjustment method stated explicitly. Unadjusted series retained alongside, because which one is correct depends on what is being tested.
Crypto tick archives
PlannedCrypto trades continuously and its microstructure is not the same as an equity session. Tick-level data is what makes a realistic fill model possible for strategies whose edge is measured in basis points.
Venue-specific. A fill model calibrated on one exchange's book does not transfer to another without re-calibration, and results will be reported per venue rather than pooled.
Macro series
LiveRegime detection needs macro state. Most published macro backtests leak because they use the current revised value of a series rather than the value that was actually printed at the time.
Published now — see the macro page. Every series carries its observation date and a freshness verdict judged against its own release cadence, not a flat clock. Vintage-aware history (ALFRED first prints, rather than today's revised values) is NOT yet implemented: the current artifact serves latest-revision values, which is fine for monitoring and wrong for backtesting a strategy that would have seen the first print. That gap is stated here rather than glossed.
Fundamentals & filings
PlannedFundamental factors are only testable if the data is keyed to the filing date rather than the fiscal period it describes. A quarter ending in March is not knowable in March.
Indexed by filing timestamp, not period end, specifically to prevent the most common look-ahead error in fundamental backtesting.
Principles
Three rules govern anything that enters the catalog, and they are the same rules that make the backtesting doctrine enforceable rather than aspirational.
- Point-in-time or it does not count. Every series must be queryable as of a past date, returning what was known then — not what is known now. Restatements and revisions are kept, not overwritten.
- Dead things stay in. Delisted tickers, defunct issuers, and discontinued series remain in the universe. Removing them is the mechanism by which survivorship bias enters, and it is usually invisible in the result.
- Adjustments are declared. Split and dividend handling, currency conversion, and any cleaning applied to raw vendor data are stated with the data rather than baked in silently.
Access
Dataset access is included from the Pro tier upward, delivered through the engine API, with each set becoming available as it is released. Neither the API nor any purchasable tier exists yet — see the waitlist to be told when it does. We will not redistribute third-party data in violation of the licence it came under, so some sources may end up available only as derived features rather than as raw series.