Calibrated forecasts that become constrained decisions.
The Forecaster predicts whether a commercial outcome will occur within a defined horizon and estimates time-to-outcome where support exists. The Decision Engine turns that forecast into a transparent action the business can accept, modify, reject, ignore, and later audit.
A forecast is not useful unless the business can trust its limits.
Levered must never express more certainty than the data supports. Each forecast is tied to a timestamp, outcome definition, horizon, segment support, calibration standard, model version, and decision rule before it reaches the action queue.
The point in time when the forecast is made and the decision must be valid.
The commercial event being predicted, such as conversion, progression, purchase, revenue or agent action.
The window for the prediction: 7-day campaign conversion, 14-day account progression, or another defined period.
Effective segment n and data quality at the recommendation granularity.
Brier score, calibration slope/intercept, expected calibration error and baseline comparison.
The transparent rule that turns the forecast into scale, cut, hold, wait, retest, route, approve, challenge, reject or insufficient evidence.
One table can hold investor, RevOps, product, GTM and agent-trust decisions because each row uses the same contract: forecast, support, action, outcome, audit.
| ID | Owner | Prediction | Horizon | Forecast | Support | Decision | Audit |
|---|---|---|---|---|---|---|---|
| INV-04 | Investor | Net revenue retention risk | 30d | 37% · 34-41% | n=1284 | PRIORITISE | observed |
| REV-12 | RevOps | Account progression | 14d | 51% · 46-56% | n=842 | ROUTE | eligible |
| PM-09 | Product | Activation lift | 21d | 44% · 38-49% | n=617 | HOLD | observed |
| GTM-31 | GTM | Segment conversion | 7d | 39% · 35-43% | n=1296 | SCALE | eligible |
| AGT-18 | Agent trust | Request outcome | session | 62% · rules | n=184 | CHALLENGE | policy |
| MKT-07 | Marketing | Creative fatigue | 10d | 28% · 20-36% | n=311 | HOLD | not yet |
▲ DECISION · route REV-12 because 14-day progression forecast is 46-56%, support is n=842, and outcome logging is audit-eligible.
A sophisticated model is not enough. The alpha standard is whether the forecast is calibrated on a time-based holdout, compares honestly against naive baselines, and improves the decision without hiding uncertainty.
From prediction unit to decision record.
The Forecaster does not ship because it is statistically sophisticated. It ships only when it remains calibrated, beats or matches agreed baselines, and produces a decision record operators can inspect.
Choose the commercial question, timestamp, outcome label, horizon and segment granularity before modelling.
Confirm labels, timestamps, censoring, non-converters, source rights, and effective segment support.
Compare against historical segment rate, recency-weighted average and simple rule-based models.
Widen intervals, back off to broader segments, or return insufficient evidence when support is weak.
Map the forecast and constraints into one inspectable action with reason codes and expiry.
Log user response, observed outcome and audit linkage so historical forecasts remain reproducible.
Prediction families for alpha
Each family starts with a clear question, preserved non-converters, time-aware splits, outcome labels, and data rights. Agent outcomes remain rules-based until enough labelled request-to-outcome pairs exist.
Will this geo, audience, channel or segment convert within a defined horizon?
Initial method: discrete-time hazard or gradient-boosted survival model.
Will this lead or account progress within the next 14 days?
Initial method: survival model with behavioural and CRM features.
Is this verified request likely to produce a commercial action?
Alpha method: rules-based until labelled data supports training.
Strong support shows a point forecast and interval. Moderate support restricts action strength.
Weak support backs off to a broader segment or returns insufficient evidence.
If data drift is detected, confident decisions are suspended for affected segments.
The system flags recalibration instead of pretending certainty survived.
Forecast logs store inputs, outputs, support, model version and explanation.
Historical forecasts must be exactly reproducible from versioned inputs.
goal_id entity_id // campaign, account, product, region, agent forecast // probability · horizon · interval segment_n model_version decision_type // scale · cut · hold · wait · retest · route · approve recommended_delta expected_impact reason_codes[] constraints[] expires_at user_response // accepted · rejected · modified · ignored observed_outcome causal_audit_id
Scale demand only when the expected revenue does not destroy contribution margin.
Move quickly where evidence is strong; retest when the signal is interesting but not yet trustworthy.
Permit agent actions when the scope is narrow and the graph can explain the commercial reason.
Do not create demand that sales, stock, fulfilment, or support cannot absorb.
What a decision looks like
The output is designed to be readable by operators and defensible to leadership: forecast, interval, support, freshness, rationale, expected value, downside, action, expiry and audit path.
SCALE - Increase C-104 by +£800/day because London demand is pacing ahead, OpenLift shows Tier 3 incremental lift, CAC remains under target, and the revenue gap is widening.
CUT - Reduce C-160 by -£420/day. Spend is rising, contribution margin is weak, and the best available evidence suggests the conversions are not incremental.
ROUTE - Send account A-442 to sales today. Pricing intent, product usage, account fit, and return frequency have crossed the readiness threshold.
HOLD - Hold discounting on Product B. Demand is rising without incentive, stock cover is low, and margin would fall below the goal constraint.
RETEST - Run a 14-day Singapore geo test. Current lift is promising but underpowered, so the engine recommends more evidence before scaling.
APPROVE - Approve the agent report request. The principal is verified, scope is read-only, and the report supports an active revenue goal.
Fewer opinions. Faster operating decisions.
Growth teams see which action is most likely to move the target today, not another retrospective report.
Finance gets expected value, downside, support and uncertainty before a recommendation changes spend.
Sales and RevOps get account routing based on readiness, fit, evidence, and capacity.
Commerce teams can balance demand generation with margin, stock, and regional performance.
AI agents can operate against explicit goals without bypassing human-defined boundaries.