Forecasts

Forecasts are predictions with evidence and review. Operator workflow in Get your first forecast.

Forecast kinds

Kind	Answers	Example
Probability	Likelihood over event + horizon	"70% renewal risk for Acme next quarter"
Categorical	Label, now	Ticket topic "billing"
Numeric	Number with horizon	Expected order volume next month
Scenario	Conditional futures	"If hiring freezes, attrition hits 9%"

Predictor fields

Field	Meaning
Forecaster	Predictor by id or slug — one must be named
Subject	Decision target: value, record reference, or label
Horizon	Aim point + window end — no horizon, no test date
Parameters	Forecaster-specific settings

Execution modes

Mode	How it runs	Prefer when
Single call	One bounded pass	Inputs are clear — cheapest to reason about
Chain	Ordered steps (aggregate, model, ontology query, remote model, tool) ending final	Staged reasoning with inspectable intermediates
Agentic	Bounded multi-step with tools, max ~12 steps	Path cannot be scripted; watch what it did

{
  "mode": "chain",
  "steps": [
    { "kind": "aggregate", "params": { "by": "account_id" } },
    { "kind": "ontology_query", "params": { "type": "Account" } },
    { "kind": "model", "tool": "renewal-risk-v3", "final": true }
  ],
  "budget": { "maxDurationMs": 60000, "maxExternalCalls": 6 }
}

Budgets

Cap	Guards against
Max duration	Runaway execution
Max tokens / cost	Surprise spend
Max external calls	Dependency storms
Allowed hosts / tools	Exfiltration and scope creep

Budgets are guardrails, not tuning knobs — a forecast needing bigger budgets needs a narrower question first.

Image: predictor configuration showing mode, budget caps, and allowed tools.

Outcomes

State	Meaning	Action
Occurred / did not occur	Event happened or didn't	Normal close
Unknown	Horizon hasn't arrived	Leave open, don't force
Ambiguous	Evidence conflicts	Record the conflict
Unresolvable	No evidence can settle it	Close as such, stop carrying it

Example: renewal forecast closed as occurred when churn landed. Example 2: scenario marked unresolvable after its premise changed.

Choosing well

Likelihoods for risks, labels for routing, numbers for capacity, scenarios for decisions with options.
Single-call first; chains for staged reasoning; agentic only for open investigation.
Close every forecast — unmeasured accuracy is assumed accuracy.