One forecast.
Many financial questions.

YC vertical-model bar

01 — What changed

One forecast. Many questions. One sampling inference. No per-query classifier training.

02 — Architecture evidence

This proves interface and execution behavior, not predictive edge.

Queries still compute reductions over stored paths. Return thresholds concern terminal prices, not first-touch events. Drawdowns use daily closes, not intraday lows. Volatility compares sample standard deviation of log returns with the prior 20 sessions. The natural-language interpreter supports bounded phrasing and rejects unsupported requests.

03 — Does it actually predict well?

Forecast origins 2026-07-02 to 2026-10-01; primary outcomes through 2026-10-08. 1157 forecasts; 8099 total decisions across 7 query definitions. These are dependent decisions, not independent observations or prospective predictions.

All eligible local symbol files; fixed 20-session stride from row 252; origins strictly after 2026-06-30; full 20-session outcomes required; 253-row trailing context; seed 1701. No tuning. Baselines compare only positive 5-session returns on identical rows. Historical 5d bootstrap uses only trailing context. Cached AAPL replaces unavailable SPY.

Best baseline by test Brier (descriptive selection): historical_5d_bootstrap. Paired Finio-minus-best Brier difference 95% interval: -0.003283 to 0.003301. Negative would favor Finio. No clear advantage established.

1000 seeded percentile bootstrap replicates clustered by origin date; exploratory: few dates, cross-date dependence remains. Best baseline selected descriptively on test Brier, not independently validated.

Log loss penalizes confident errors (natural log, probability clipped at 1e-12). ECE is equal-width 10-bin calibration error. Lower Brier, log loss and ECE are better; higher accuracy and AUROC are better. Accuracy uses strictly >50%; AUROC is unavailable for single-class samples. Only Brier has an uncertainty interval here; point estimates for other metrics do not establish significance.

External foundation-model comparison: NOT YET MEASURED.

04 — Generalization

NVDA excluded from specialist baseline fit; forecast uses NVDA trailing context. No globally trained forecast model, so this is not pretrained zero-shot transfer.

Group / sector holdout: NOT NEWLY MEASURED. No comprehensive sector metadata. Original energy/XOM partition remains in archived study.

Query support: return_bucket between [-0.02,0.04] · 10 sessions; N=1157, Brier=0.230258. No query trains the forecast engine. Range queries reduce existing paths; this is architecture support, not pretrained task transfer.

ZERO-SHOT PRETRAINED TRANSFER: Not yet demonstrated. Few-shot adaptation of a pretrained model: Not yet demonstrated. No pretraining took place.

05 — Compute and forecast reuse

Local in-process wall time; includes initial cold forecast; no network, API, persistence or browser overhead. Query timings include mixed supported tasks. Not a cost comparison with another model.

Forecast generated once; subsequent questions reuse it. This is measured reuse within Finio, not a compute win against a specialist or external foundation model.

06 — Prospective and verifiable evidence

Created and timestamped before the evaluation outcome was available. Historical evaluations and test fixtures are excluded.

No durable v0.2 ledger export or local database found in this checkout; production database not queried.

Existing immutable D1 predictions reference forecast_id, model hash/version, cutoff, typed query, probability, horizon and protocol. Resolutions append outcome/date/Brier/log loss separately. No historical records rewritten. Production counts remain unknown without an audited export.

An immutable append-only ledger and resolver are implemented and tested. This proves recording capability, not measured prospective performance. Same-day query flags alone do not prove issuance before market outcomes; an audited export must verify issue timestamps against session completion and outcome dates.

07 — Data rights and provenance

08 — Reproducibility and limitations

Per-forecast context and outcome dates, identities, probabilities, labels and same-task baseline outputs are published in the derived evaluation rows. Source-file hashes and exclusions are in the evidence manifest. No raw cached bars are redistributed; full price-to-forecast reproduction requires separately authorized copies of those exact private files. The published rows support independent score recomputation. Timings vary by machine.

• Current architecture prototype is not a trained finance foundation model.
• No predictive edge established; losing baselines retained.
• Many decisions share a forecast; assets and dates are dependent. Confidence intervals are exploratory.
• Expanded study is a new sample; not directly comparable to v0.1 or the original 112-decision v0.2 study.
• AAPL benchmark is not SPY. SPY outcomes unmeasured locally.
• Cached data rights, acquisition timing, revisions and survivorship limit validity.
• No external foundation-model comparison or accuracy/compute Pareto win.
• Production prospective counts unavailable locally; zero verified local records is not proof of zero production predictions.