Evidence, not instinct.
Good research is a sequence of disciplined choices: what to measure, what to model, what to challenge, and what we can honestly conclude.
Every conclusion has a chain of evidence.
Our process is designed to make the links between a decision, its evidence, and its uncertainty visible.
Step 1 of 5
Frame
Turn an uncertain decision into a testable question.
We begin with the decision, not the dataset. In a scoping workshop, we surface competing hypotheses and identify the evidence that could change the answer. We agree on success measures, constraints, and the boundaries of the inquiry before fieldwork begins. The result is a research plan that makes trade-offs in time and budget explicit.
What we do
- Run a decision-focused scoping workshop
- Build a hypothesis tree and define decision criteria
- Identify stakeholders, audiences, and evidence gaps
- Set success metrics and research priorities
- Agree on governance, plan, and budget
- Methods & tools
- Stakeholder interviews · Hypothesis trees · Evidence gap mapping · Research design
- What you receive
- Decision brief · Hypothesis tree · Research plan and budget
- Typical duration
- Typically 1–2 weeks
Step 2 of 5
Collect
Gather the right evidence, not simply the most available data.
Collection combines what people say with what they do. We design primary research around the populations that matter, while assessing secondary and alternative sources for relevance and coverage. Permissions, nondisclosure agreements, and data handling are established before sensitive information moves. Recruitment and instruments are checked so later analysis rests on a credible foundation.
What we do
- Design surveys, interview guides, and sampling plans
- Recruit experts, customers, donors, or field panels
- Gather transaction, web, geospatial, and filing data where appropriate
- Assess provenance, bias, and missingness
- Put consent, NDAs, and data governance in place
- Methods & tools
- Survey design · Panel sourcing · Expert interviews · Customer and donor interviews · Secondary and alternative data review
- What you receive
- Fieldwork summary · Source inventory · Documented dataset and data dictionary
- Typical duration
- Typically 2–5 weeks
Step 3 of 5
Analyze
Separate durable patterns from noise and convenient assumptions.
We clean and weight the evidence before asking a model to explain it. Analytical choices follow the question: segmentation for distinct needs, choice models for trade-offs, forecasting for change over time. Where data is sparse, we show how estimates borrow strength and where they remain uncertain. Scenarios translate model outputs into decision-relevant alternatives.
What we do
- Clean, link, and weight the data
- Build behavioral segments and test stability
- Model demand, price sensitivity, and adoption
- Forecast under explicit economic assumptions
- Compare scenarios against the decision criteria
- Methods & tools
- Clustering and segmentation · Conjoint / MaxDiff · Econometric models · Bayesian hierarchical models · Scenario analysis
- What you receive
- Analysis workbook · Segment profiles · Scenario model and assumptions
- Typical duration
- Typically 2–4 weeks
Step 4 of 5
Validate
Challenge the result under different scenarios and competing explanations.
A plausible model is not yet a defensible finding. We back-test where histories permit, reserve holdout evidence where feasible, and probe sensitivity to key assumptions. Independent review challenges the design, code, and interpretation. We report confidence ranges and limitations next to the conclusion, not in a footnote.
What we do
- Back-test forecasts and use holdout samples where feasible
- Run sensitivity and Monte Carlo scenarios
- Triangulate primary, secondary, and behavioral evidence
- Arrange independent senior or PhD-level peer review
- State confidence ranges and limitations plainly
- Methods & tools
- Holdout / back-testing · Sensitivity analysis · Monte Carlo simulation · Source triangulation · Independent peer review
- What you receive
- Validation memo · Sensitivity ranges · Reviewed assumptions and limitations
- Typical duration
- Typically 1–2 weeks
Step 5 of 5
Advise
Translate the analysis into decisions and a clear path forward.
We communicate findings in the language of the decision and the people responsible for it. The executive readout separates what is known, what remains uncertain, and what should happen next. Teams receive the models and definitions needed to revisit the reasoning rather than a static presentation alone. Follow-up can test implementation signals at 30, 60, and 90 days when appropriate.
What we do
- Write a board-ready decision memo and report
- Lead an executive readout and discussion
- Prioritize options and implementation dependencies
- Hand off models, definitions, and assumptions
- Agree on a 30/60/90-day follow-up plan
- Methods & tools
- Decision synthesis · Executive workshop · Model hand-off · Implementation planning
- What you receive
- Board-ready report · Executive readout · Decision memo · Model hand-off
- Typical duration
- Typically 1–2 weeks
Ask the right people the right questions.
Primary evidence is strongest when the sample, instrument, and interpretation are designed together.
Surveys & panels
We design instruments around the decision, not an inventory of interesting questions. Sampling plans, cognitive testing, weighting, and nonresponse checks help establish what the answers can support.
Expert interviews
Structured interviews with market participants surface mechanisms behind trends, expose weak assumptions, and sharpen later quantitative work. We distinguish expert judgment from independently verified evidence.
Customer & beneficiary research
Interviews and panels reveal how people actually make choices. We adapt recruitment and language to the audience, including respondents who are hard to reach or underserved by conventional research.
Methods chosen for the question.
Technical sophistication is useful only when it helps answer something the client actually needs to know.
Bayesian hierarchical models
Pool information across related groups without erasing differences. Particularly useful when individual segments have limited observations.
Time-series & econometric forecasting
Model demand, rates, and market change while making assumptions and prediction intervals explicit.
Causal inference
Use comparison groups and approaches such as difference-in-differences when the question is what changed because of an intervention.
Conjoint & discrete choice
Reveal trade-offs across price, features, and service levels through carefully constructed choices.
Clustering & segmentation
Identify groups with distinct behaviors or needs, then test whether the segments are stable and actionable.
Machine learning
Apply predictive methods when they improve out-of-sample performance and can be interpreted responsibly.
NLP & unstructured text
Analyze interviews, open-ended responses, and other text at scale without losing qualitative context.
More signals. Better judgment.
Transaction, web, and geospatial data can reveal behaviors conventional research misses. We assess provenance, coverage, bias, and permissions before treating a feed as evidence. New data is not automatically better data.
Rigor is a practice.
Quality is built into research design and review, not applied as a final polish.
Triangulate before concluding
We compare evidence from multiple sources and investigate meaningful disagreement rather than forcing a tidy average.
Make uncertainty legible
Confidence intervals, scenario ranges, and limitations belong alongside the headline finding.
Review the work
Senior researchers and PhD-level expertise review designs, models, and interpretations. Critical assumptions receive independent challenge.
Protect what is entrusted to us
We work under nondisclosure agreements where required, limit access to sensitive information, apply appropriate data handling and retention practices, and keep client identities confidential.
Separate correlation from causation
We reserve causal claims for research designs that can reasonably support them. When evidence is descriptive, we say so.
Leave a usable record
Our deliverables make definitions, methods, and key assumptions traceable so teams can revisit the reasoning after the presentation ends.
Research specifications.
The toolkit changes with the question. The standards do not.
- Models
- Bayesian hierarchical · Econometric forecasting · Causal inference · Discrete choice · Segmentation
- Primary data
- Surveys · Panels · Expert interviews · Customer and beneficiary research
- Alternative data
- Transaction · Web signals · Geospatial
- Quality assurance
- Design review · Assumption testing · Peer review · Sensitivity analysis
- Data handling
- Confidentiality agreements · Access controls · Appropriate retention practices
# model the signal; test the uncertainty y = Xβ + ε P(θ | data) ∝ P(data | θ) P(θ)