Forecasts the expected high-low range — not price direction — for each of the next eight 15-minute candles. The current /v2 arm is PF-WLS; /v1 preserves HAR-PF-cal as the rollback path.

Each range forecast is converted to a per-bar σ proxy (range / 2.4). Those variances are accumulated into horizon-specific offsets from 15 minutes through two hours. The cone remains centered on the latest close; it forecasts movement scale, not the future center price.

PF-WLS was promoted after 189 strictly post-enrollment live folds: mean QLIKE was 0.4% lower than HAR-PF-cal (3.00793 → 2.99575; Δ −0.01219), with one-sided DM-HAC p=0.0064.

Descriptive reconstruction from the frozen research promotion ledger, using the /v2 cone formula: 171 folds were clock-contiguous two-hour forecasts, covering 1,368 target candles. 58.4% of closes and 38.1% of full high-low candles stayed inside the published ±1σ band; at ±2σ, coverage was 89.4% of closes and 79.3% of full candles. At the two-hour endpoint, 58.5% of closes and 42.1% of full candles were inside ±1σ. The same research ledger's post-promotion slice through 2026-07-15 — not served-request logs — showed similar 59.2% close coverage across 608 clock-contiguous targets.

The production volatility estimate — consumed live by the nf cockpit's volatility cone.

At each eligible complete GC COMEX candle, forecasts the trading volume of each next 15-minute candle through two hours (eight separate targets). The response includes both model-scale log1p_volume and raw-contract volume. It does not forecast price direction.

A frozen strictly-forward research result exists, but its numerical performance and source-derived examples are withheld from the public page until source-specific derived-output permission is pinned.

Methodology remains public; fold counts, loss values, horizon results, vectors, and source hashes remain internal during the authorization hold.

Situational liquidity/activity context — whether the not-yet-complete target bars are likely thin or busy.

The technical entitlement is real-time with no fixed source delay. Automated collection and atomic-live publication are paused pending written market-data authorization. The external endpoint returns no forecast or historical-bar payload during this hold. Completed-bar publication, processing, and exchange closures would still affect anchor age after a lawful activation.

At each eligible complete XAUUSD spot-venue candle, forecasts the venue volume of each next 15-minute candle through two hours (eight separate targets). The /liquidity/v1 API serves the frozen XAUVOL-2 NBEATSx seed-0 artifact.

A frozen strictly-forward research result exists, but its numerical performance and source-derived examples are withheld from the public page until the applicable upstream derived-output permission is pinned.

Methodology remains public; fold counts, loss values, horizon results, vectors, and source hashes remain internal during the authorization review.

Expected XAU trading activity / liquidity context. The service only anchors on complete 15-minute bars and exposes target-level timing status.

The XAUUSD spot aggregate source retains a nominal 10-minute delay. The service verifies all 15 unique source minutes and volume parity before selecting an anchor; read timing.target_steps for forming or expired targets.

At each complete GC COMEX candle, forecasts the within-bar mean quoted bid-ask spread ($/oz) of each next 15-minute candle through two hours (eight targets) — i.e. how expensive it is likely to be to cross the market. It forecasts liquidity cost, not price direction.

NBEATSx (input 96 bars), future exog = deterministic calendar clock, hist exog = order-book activity (realized volatility + gross order-flow imbalance). The spread is trained as origin-anchored innovations and recomposed to level inside the model. Served as a 5-seed ensemble (per-seed clip at 0, equal-weight mean).

A frozen reserved-tail confirmation exists, but its numerical performance, source-derived examples, and vectors remain internal until source-specific derived-output permission is pinned.

Situational liquidity-cost / execution-timing context — a plain cheap / normal / expensive read on how costly crossing the GC spread is likely to be over the next two hours.