Step / 01
Claim strength
Every method is labeled descriptive, predictive or causal. The label travels with the number into every brief, so a correlation is never presented as an effect.
Platform · Measurement
Only a causal method gets to promise a result.
Measurement decides what Signal is allowed to claim. Descriptive methods size an opportunity, predictive methods forecast it, and only causal methods may promise an effect on the P&L. Every decision is registered with its control group before it starts, so nobody picks the result after the fact.
Step / 02
A contract before the decision
Each decision is pre-registered with a metric, at least two treatment and two control units, a start, an end, a readback date and a guardrail metric.
Step / 03
Readback by difference-in-differences
At least three periods before and two after, every observation used, a 95% interval and a parallel-trend check. The window is fixed, so it cannot be chosen to flatter the result.
Step / 04
Learning that waits for approval
What a readback teaches lands as a candidate rule. It stays inactive until a person approves it, so the system only learns what people agreed to.
Methods
The method behind each claim.
MethodWhen it is usedClaim strength
Difference-in-differences
Reading back a decision against its pre-registered control group.
Causal
Synthetic control
When no clean control group exists, a weighted combination of similar units stands in.
Causal
Monte Carlo simulation
Ranges for a scenario before it is approved.
Predictive
Forecast backtest
Every forecast is tested on held-out history against a naive baseline.
Predictive
Price elasticity
Sizing how volume responds to price, per SKU.
Descriptive
In development: independent scientific validation of readbacks, and accounting recognition of captured value. Until then a readback is reported as a measured effect pending validation, never booked as captured value.
A decision in. An effect you can defend out.
It reads
- 01The decision and its alternativesDoing nothing is always one of them
- 02Treatment and control unitsFixed before the start
- 03Periods before and afterAt least three and two
- 04A guardrail metricWhat must not get worse
It returns
- 01The effect with a 95% intervalOr a plain statement that it was not measurable
- 02A parallel-trend checkWhether the comparison holds
- 03A claim-strength labelDescriptive, predictive or causal
- 04A candidate ruleFor a person to approve
From the field
All cases →
Where this has been used.
Data everywhere, decisions nowhere.
R$22–42Mgross profit a year at national scale
R$4.4–8.55Mone-time cash from stock parked at the distributor
97%of revenue covered by the rebuilt mapping
60% vs 32%forecast error, official against a naive average
Media decided on 10% of the sales.
232%of sales credited to one channel by the old attribution
60%of the media budget below an ROI of 1
+15%digital sales from reallocating the same budget, model estimate
Measure what
moved
Every Value Tree names the decision, the metric and the control group before any work starts.