Tauric Research
Perspectives

Data foundations

  • Foundation

Point-in-time data, provenance, and why training cutoffs matter in finance.

A language model trained on recent web text has read how past earnings seasons, rate decisions, and crises turned out. When it is asked to judge a period it has already read about, the result mixes judgment with memory. Bridgewater’s AIA Labs, reporting on an AI forecasting system, note that events before a model’s knowledge cutoff “are likely to have been memorized by the LLM … making retrospective analysis difficult” (Alur et al., 2025 (external site)).

State of the field

The usual response is to control what a system can see at each point in time, and it is harder than it sounds. The same report calls reliably enforcing an information cutoff “a surprisingly challenging problem”: search providers can leak news of events that have already happened, and “in the extreme case, this bias can create the illusion of forecasting skill where none exists”. In their own system they estimated that about 1.65% of search results carried some foreknowledge, and found that it did not materially change the results.

Practitioners build the same discipline into the workflow. Man Group’s research assistant drafts a plan in which each step declares the tool it will call and the output it expects, “incorporating safeguards to ensure data integrity and prevent lookahead bias”, and the plan goes to the user for approval before it runs (Luk et al., 2025 (external site)). In FinanceHarness (external site) (Xiao et al., 2026), every document in the search corpus carries a publication date, documents published after a question’s cutoff are blocked, and a run with detected leakage of future information is treated as invalid rather than merely scored lower. The same care applies to structured data: a revised economic series, a restated financial statement, and a security that later changed its identifier each show today a version of the past that was not available at the time.

Principles

A figure without a date is incomplete. Research data should say what was known, when it was known, and where it came from, and an analysis should be able to name every source it used. In finance, a result that cannot be checked should not be relied on.

Our approach

We serve market, reference, filing, and news data from one source of record, with point-in-time reads, through the Tauric API and MCP server. Our aim is that analysts and agents work from the same ground truth, and that every result traces back to its inputs.

References