Engine: Deterministic (no LLM)
1. What it does
Projects a few periods ahead for a series on the dashboard, with a likely range instead of a bare number, and tells you how much this projection usually misses on your own history.
2. Deterministic or LLM?
100% deterministic. There is no LLM. It's a linear regression (trend line) with a range derived from the residual error, plus a walk-forward backtest that measures quality. No I/O, reproducible.
Parameters:
- Minimum points to project: 6 (
FORECAST_MIN_POINTS) — below that, it returnsnull(no projection). - Maximum projected periods: 24.
- The range (~95%) widens the further it gets from the last data point — projecting 12 months doesn't carry the same confidence as projecting 1.
3. Data sent and egress
None. It runs on our server over the snapshot. Nothing leaves the product, no tokens.
4. Privacy and retention
There is no egress. The quality evaluation (walk-forward) uses only the series' own history, with no leakage from the future into the fit — and there's a test proving it.
5. Tokens, cost, and budget
Not applicable — deterministic, zero AI cost.
6. How to turn it on, who can use it, and how to turn it off
A dashboard visualization feature (per widget). It doesn't go through the external-sending and budget checks.
7. Refusals
- Not enough history to project (< 6 points) → doesn't project (returns
null). - Not enough history to evaluate the forecast → verdict
insuficienteand no quality number is shown: "A MAPE over 2 points would convince exactly the people who shouldn't be convinced."
8. Limitations and absence of causality
- It's a simple linear regression: it does not capture seasonality, cycles, or regime
changes. The limitations are shown alongside every forecast, including a "reliable" one
(
LIMITACOES_PREVISAO). - The evaluation uses the series' own history (backtest) — past performance doesn't guarantee the future.
- Projecting a trend is not claiming a cause: the line says "if it keeps going as it has," not "this will happen" or "X causes Y."
- The verdict is honest —
confiavel(reliable) only when the error is small (MAPE ≤ 15%, when measurable) and the range delivers on its promise (≥ 80% of actual values fall inside it); otherwise,instavel(unstable).
9. How to validate
- Reproducibility and no leakage: the projection is deterministic and the walk-forward backtest doesn't see the future (the same series always produces the same projection and the same quality).
- No numbers when insufficient: short series → verdict
insuficiente, with no quality metric. - Range grows with distance: check that the likely range widens for more distant periods.
- No egress: no external calls, no
ai.usageevent.