Point-in-Time Language Models

Language models that know only what was knowable at a given date — and the research that builds, tests, and uses them.

A conventional language model is trained on a temporally indiscriminate pile of text. Ask it to “forecast” 2015 from 2014's news and it quietly cheats: it has already read 2016. A point-in-time (or chronologically consistent) language model is trained only on text available up to a fixed cut-off date. A sequence of such models, one per year, quarter, or month, lets you replay history without look-ahead bias.

The idea began in finance, where leakage invalidates a backtest, but the same construction is a time capsule for any field in which knowledge changes: medicine, law, journalism, policy, and the social sciences. This site collects the model sequences that exist, the corpora one could train them on, the literature on look-ahead bias and temporal generalisation, and plans for what to build next.

dated text cut-off t₁ cut-off t₂ cut-off t₃ model M(t₁) trained on text < t₁ model M(t₂) trained on text < t₂ model M(t₃) trained on text < t₃ forecast (t₁, t₂] forecast (t₂, t₃] forecast (t₃, ·] each model sees only its past; the backtest has no look-ahead by construction

Cut the dated stream at t₁ < t₂ < t₃, train one model per cut, and evaluate each model only on the interval after its cut-off. Leakage is removed structurally, not by prompting.

Model sequences

Six open efforts have released sequences of date-stamped checkpoints. They differ in domain, cadence, and in whether the first checkpoint was itself trained from scratch on period text or adapted from a modern base model (which reintroduces some leakage). Details, checkpoint lists, and a coverage chart are on the models page.

FamilyArchitectureCoverageCadenceCorpus
ChronoBERT / ChronoGPTBERT-style encoder; GPT-style decoder; instruct variant1999–2024yearly (26 + 26 + 26)chronologically filtered web and news text
Scaling PiT LMsdecoder-only, up to 4B params2013–2024monthly1T chronologically filtered FineWeb tokens
Time Machine GPTGPT-22011–2022yearly (12)WMT News Crawl + Wikipedia, datasets released
TimeLMsRoBERTa-base (and one large)2019-Q4–2022-Q4quarterlyTwitter
StoriesLMBERT-style, expanding windows1900–1963yearly (64)American Stories newspapers
HistBERTBERT-base, continued pretraining1900s–2000sdecadal (10)COHA

Research threads

The literature clusters into a few threads. The literature map draws the connections; the timeline dates them.

Building the models

Train a sequence of checkpoints on chronologically filtered text. Recent work shows the performance gap to unconstrained models narrows with scale.

He et al. 2025; Kelly et al. 2026; Drinkall et al. 2024; Loureiro et al. 2022

Measuring look-ahead bias

How much of an LLM's “forecasting skill” is memorised outcome? Detection statistics, fake-date tests, and finance-specific benchmarks.

Glasserman & Lin 2023; Sarkar & Vafa 2024; Gao et al. 2025; Benhenda 2026

Can prompting substitute?

Telling a model to “pretend it is 2015” is cheap. It works for direct queries and fails for causally downstream knowledge; models struggle with chronology itself.

Gao et al. 2025; Asai et al. 2026; Wongchamcharoen & Glasserman 2025

Temporal generalisation

The NLP thread that preceded finance: models degrade as the world moves past their training window, and continual pretraining only partly helps.

Lazaridou et al. 2021; Luu et al. 2021; Dhingra et al. 2021; Jang et al. 2022

Temporal knowledge benchmarks

What does a model know as of when? Cutoff tracing, chronological knowledge across domains, evolving medical guidelines.

Cheng et al. 2024; Park et al. 2024; Guan et al. 2026

Beyond finance

Historical LLMs as simulated informants for behavioural science, and proposed uses in medicine, law, history, journalism, and policy.

Varnum et al. 2024; the essay on this site

Corpora

Every checkpoint sequence is only as good as its dated text. The corpora page catalogues candidate sources by domain, with coverage, licensing, and the finest temporal granularity each supports, from Chronicling America and COHA to SEC EDGAR, Hansard, and Common Crawl derivatives such as FineWeb.

What to build next

The plans page holds three costed proposals for extending the open model and corpus sequences on a small budget: a focused $500k programme, an open-source corpus programme, and a decentralised Bittensor-subnet variant. They are working documents, not commitments.

Further reading

Bibliography

Grouped by thread. Every entry also appears on the literature map and the timeline. arXiv identifiers were checked against the arXiv API; SSRN and journal items against Crossref.

Model sequences

Look-ahead bias and training leakage

Prompted cut-offs and chronology

Temporal generalisation and misalignment

Temporal knowledge benchmarks

Corpora and data

Beyond finance

Working on a point-in-time model, corpus, or benchmark that is missing here? Open an issue or pull request on microprediction/pitllm.