How Cambent forecasts the TKX
The TKX forecast projects where the TKX may go. It is a range, not a promise. It targets economic capability cost — what a unit of standardized AI capability costs after quality is accounted for — not any provider's list price.
Rising means cheaper
A rising TKX means a unit of AI capability is getting cheaper versus the 2026-08-04 baseline of 100. That is the direction we treat as good. A falling TKX means capability is getting more expensive versus that baseline.
Up on the TKX means cheaper capability versus the 2026-08-04 baseline of 100. That is good.
Experimental or citeable
Every forecast run is in exactly one state: Experimental or Citeable. The state is computed from the run's evidence. There is no manual "promote to citeable."
Experimental (default at launch, and whenever the citeable bar is not met)
- Lives on the Forecast tab only, labeled Experimental, with uncertainty bands shown.
- Not presented as an official Cambent number.
- Not placed in quiet or citeable board copy, the quiet API, press, or any "Source: Cambent TKX" citation.
- Language: projected, model-implied, experimental estimate — always with the range.
Citeable (earned only)
- Reached only when every citeable-bar criterion below passes for that run.
- May appear on the Forecast tab without the Experimental label, and — only then — in citeable / API / citation contexts as an official Cambent forecast.
- Still range-first: still a projection with an interval.
What earns citeable
A forecast is citeable only when all of the following are true for that run. If any fails, the forecast stays Experimental.
- Sound index — the underlying index methodology is the locked, verified recipe the forecast trains on.
- Enough comparable history — enough history on the current methodology version (no un-spanned structural break). Starting default: about 90 days; Data may tighten or loosen via the improvement rhythm without changing this honesty rule.
- Backtest beats naive — the model beats simple last-value and drift baselines out of sample on the target horizon.
- Calibrated intervals — the bands cover outcomes about as often as they claim (starting default: within about ±10 percentage points of nominal).
- Fresh close — the run used a finalized official daily close no older than the freshness default (starting default: 2 days).
- Confirmed series only — the forecast trains on the confirmed TKX series, never a provisional market-impact series.
Citeable is earned by the data on each run. There is no button to mark a forecast citeable by hand.
Intervals and bands
- A projection is always shown with its uncertainty band.
- Allowed words: projected, model-implied, estimate, range, band, interval.
- Even in Citeable state, the band stays. The midpoint is not a locked price.
Boundary handling
What you see when something goes wrong.
| Situation | What the product does |
|---|---|
| Forecast job fails or errors | Keep the last successful forecast. Show it is stale (e.g. Last updated [date]). Never publish a partial or broken run. |
| Official daily close missing or not finalized | Do not run a new forecast. Keep the prior. Show stale. |
| Not enough comparable history | Forecast tab says insufficient comparable history. Stay Experimental. Never invent a precise number to fill the gap. |
| Methodology break not spanned | Do not silently regress across the break. Stay or drop to Experimental if comparable history is too short. Annotate the transition when shown. |
| Unusually large jump vs prior run | Hold for review; prior forecast stays current until cleared. Never auto-suppress without indicator. Never silently replace. |
| Interval calibration falls below the citeable bar | Automatically drop Citeable → Experimental (and off the citeable board) until the bar is met again. State is never sticky. |
Never:
- Partial or errored forecast presented as official
- Experimental forecast on the quiet citeable board or in "Source: Cambent TKX"
- Point forecast without its band
- Manual promote-to-citeable
Limitations
- The forecast is a model of capability-adjusted index levels, not a vendor price sheet and not your invoice.
- Bands describe uncertainty; they are not a guarantee of future prints.
- Experimental forecasts are for exploration on the Forecast tab.
- Citeable forecasts are still projections with intervals — official, not certain.
- Cambent does not send your traffic. The forecast does not route models.
The live citeable index (today's confirmed TKX, Open, and Frontier with as-of) stays on the quiet board. A forecast joins that citeable surface only when it is Citeable.