What decision changes?
Name the choice, who makes it, its constraints, and the business outcome: margin, retention, risk, speed, or qualified demand.
← Soft MeasuresA problem we work through · 03
“We don't know what drives our best results.”
Turn observed results into decisions you can test.
A strong result may reflect skill, timing, selection, or chance. We connect the data behind it, define the decision it should inform, and test whether the pattern survives new cases. The output is an operating rule, model, or recommendation whose evidence and limits are visible.
First, understand the situation
What we would look at.
Can we trust the inputs?
Join events, transactions, customer or market context, and outcomes. Check missing records, changing definitions, delayed outcomes, and information that would not have been known at decision time.
Does it beat the current rule?
Compare a proposed score, forecast, or policy with the existing process on historical and held-out cases. Examine errors and the cost of acting on a wrong answer.
Then, build around the answer
How the work takes shape.
An auditable data foundation
Ingest, normalize, and reconcile source data; keep definitions, provenance, and uncertainty attached to the result.
A test matched to the decision
Use rules, statistical analysis, scoring, forecasting, simulation, or machine learning only where each adds measurable value.
A guarded decision workflow
Present the recommendation with its evidence and limits, require human approval where consequences are material, and measure what happened afterward.
Inside the decision system
A model is useful only when its decisions hold up.
Technical depth matters most in the evaluation: what the system knew at the time, what it would have recommended, and what happened when that recommendation met reality.
Define the decision contract
Specify the action, available inputs, success measure, timing, and hard constraints. This prevents a model from optimizing a convenient metric that does not improve the business choice.
Establish a baseline
Reconstruct the current rule or expert process and measure its outcomes. Separate training from later or held-out cases, and check for data leakage, segment differences, and missing information.
Test before exposure
Compare candidate rules or models using backtests, simulation, sensitivity checks, and error analysis appropriate to the domain. Record assumptions and where the evidence is too thin for a recommendation.
Monitor the live choice
Put approval and limits around consequential actions. Log inputs, recommendation, decision, and eventual outcome so the system can detect drift, explain failures, and be revised from evidence.
In our work
Examples of the thinking in practice.
Quant · Research
RLTrade
Market-data pipelines, reproducible experiments, held-out evaluation, simulation, risk limits, monitoring, and human approval separate a research result from a live trade.
Explore product →Fintech · M&ABizValuePro
A deterministic SDE and EBITDA valuation engine combines business inputs with comparable-company data and explicit confidence handling before a value is presented.
Explore product →Your situation
Let's look at the real problem.
Bring us a result you want to understand or a decision you wish you could test before committing to it.