dragonlimited/DragonData-DatedDragon

DragonData DatedDragon (ROADMAP)

One model per year — guaranteed never to have seen the future. Annual-cutoff variants of the DragonData finance line, each trained strictly on documents dated at or before December 31 of its year. The honest substrate for backtesting and finance research: no lookahead bias, verified by perplexity-reversal checks.

Status: PLANNED (S8 — starts after flagship + align). Public roadmap so the method is auditable before the first cutoff trains.

Why it matters

Every internet-trained model has seen the future, silently invalidating backtests. DatedDragon bounds knowledge by construction: ask the 2019 model about 2020 events and it must not know. Verification: perplexity reversal after each cutoff date.

Method (locked)

• Cutoff = document date ≤ Dec 31 of model year (filings, news, prices all timestamped).
• Annual models (e.g. dated-2019 … dated-2025) as revisions/folders in this repo.
• Each ships: weights + cutoff certificate (perplexity-reversal report) + exact data manifest.

Where to find things (lands at S8)

├── README.md ├── dated-YYYY/ (weights + certificate + manifest each) └── EVAL.md

Family: FinDense-3B · Corpus · CN-FinEval.

License

Apache 2.0.