- Home
- Technology
- Data Science
Discipline
The statistics behind the claim.
Most questions put to a data science team are causal — will this change work? — and most answers given are correlational. The distance between those two things is where a lot of money gets lost.
01 / What this covers
Experiment design and causal inference.
Data science here means the part that comes before and after modelling: framing a question so it can be answered, designing a study that can actually answer it, and interpreting the result without overclaiming.
It is the least glamorous and most frequently skipped discipline in applied AI, and it is what separates 'the metric went up' from 'the change caused the metric to go up'.
02 / In practice
What this looks like in a build.
- Experiment design
- A/B and switchback tests with power calculated in advance, so you know before starting whether the test can detect an effect worth acting on. Underpowered tests are worse than no test — they produce a confident null and stop good ideas.
- Quasi-experimental methods
- Difference-in-differences, regression discontinuity and synthetic controls for the many situations where randomisation is impossible — a pricing change, a regional rollout, a regulatory shift.
- Uncertainty quantification
- Intervals on every estimate, and an explicit statement of what would have to be true for the conclusion to be wrong. Point estimates presented without a range invite decisions the evidence does not support.
- Metric design
- Choosing measures that are hard to game and correspond to something the business actually values. Any metric that becomes a target will be optimised, including in ways nobody intended.
- Analysis that is reproducible
- Version-controlled, reviewed, and runnable by someone else. An analysis that exists only in a notebook on one laptop is not evidence anyone can rely on.
03 / Honest limits
Where we would tell you not to.
Observational data supports causal claims only under assumptions, and those assumptions are frequently untestable. We state them explicitly rather than burying them, which occasionally makes a result look weaker than a less careful analysis would have.
Statistical significance is also not business significance. A real effect can be far too small to justify the cost of acting on it, and we will say so.
04 / Related
The rest of the method.
These disciplines are applied together far more often than separately. Most engagements draw on all three.
Deep Learning
Neural architectures, fine-tuning and the compute discipline they demand.
Robotics & Autonomy
Perception, planning and control for machines that move.
Technology overview
The engineering standards that apply across every discipline — evaluation, reproducibility, failure behaviour and cost control.
Next step
Have a technical question about this?
Technical enquiries reach an engineer rather than a sales inbox. If you want to interrogate the approach before committing to anything, that is a conversation we would rather have early.