Formative

Numeric decision support, for when you can't A/B test.

Join the pilot

We use your email address and message only to reply about the pilot, and never share or sell them. Under the GDPR you can ask us to see, correct or delete them at any time at contact@liminalplanes.com.

The conversation on the shops example: every approach's estimate of what +1 price does to sales, the verdict that 3 of 3 agree price lowers sales, and an offer to weigh it as a decision.
The canvas's Estimates tab beside it: backdoor adjustment, interrupted time series and difference-in-differences each estimate price lowers sales, with their intervals, against the unadjusted slope that says it raises them.
Fig. 1 The app on simulated data with a known effect of −4: (a) the conversation, (b) the canvas beside it. Three approaches resting on different assumptions recover the effect; the unadjusted slope, confounded by catchment income, has the wrong sign.

Decision-making in natural language.

Specify your question in natural language and work through the problem, or use the built-in canvas for larger edits. Both update each other.

  1. The Data tab: the shops table locked, with its columns to model.

    Step 1

    Data

    Start from your data. Check the columns to model, mark rows from experiments, and lock it.

  2. The conversation, asking what you want to know about the table.

    Step 2

    Question

    Ask your causal question in plain words. Formative's autodiscovery drafts a causal graph, based on your data.

  3. The Graph tab: the causal graph, approved and locked.

    Step 3

    Graph

    Refine the graph based on expert opinion, and lock it.

  4. The Decide tab: the recommendation not to increase price, and how far it holds.

    Step 4

    Estimate

    Estimate cause and effect using state of the art methods and refutations derived from your graph.

Every answer rests on a graph.

All causal estimates rely on assumptions. Some are testable, some are not, but they always exist. Formative treats assumptions as first-class citizens: they are never hidden, and are saved as part of any estimate.

The brief for the shops example: the question, the graph approved and locked, the verdict that price lowers sales with 3 of 3 approaches agreeing, and what it rests on, each item marked confirmed, judgement or assumed.
Fig. 2 The brief on the shops example: the question, the verdict, and everything it rests on.

Built by those who care.

Formative is built by a data scientist and programmer who has done this work outside the textbook, and seen much of what goes into decision making based on facts and not feelings.

No estimate is perfect, indeed some may be wrong. The bar to clear should be a marked improvement on how businesses currently operate.

Have a question you couldn't answer with an A/B test? Bring it, with the table you have.

Join the pilot

We use your email address and message only to reply about the pilot, and never share or sell them. Under the GDPR you can ask us to see, correct or delete them at any time at contact@liminalplanes.com.