On this page

Distribution Reference

Every int/float column carries a distribution in its distribution_params. Pick the shape that matches the real-world quantity; pass the listed parameters. All distributions honour min/max clamping and decimals.

Python
{"type": "integer", "distribution": "poisson", "lambda": 4, "min": 0}

Continuous#

DistributionParamsUse for
normalmean, stdSymmetric quantities: age, height, test scores.
lognormalmu, sigmaRight-skewed money: price, salary, order amount.
uniformmin, maxFlat ranges: latitude, a 1–5 rating.
exponentialscaleWait times, inter-arrival gaps.
betaa, b (scaled to min/max)Bounded proportions, scores in a fixed band.
gammashape, scalePositive skewed durations, insurance claim sizes.

Discrete#

DistributionParamsUse for
poissonlambda (alias lam)Counts per interval: items per order, calls per hour.
binomialn, pSuccesses out of n trials: conversions, defects.

Heavy-tailed (power-law)#

For "a few get most, most get very few" — views, followers, wealth, file sizes:

DistributionParamsNotes
zipfaDiscrete power-law; larger a ⇒ steeper tail.
pareto / power_lawalpha, scaleContinuous heavy tail.
Python
{"type": "integer", "distribution": "zipf", "a": 2.0, "min": 1}     # view counts
{"type": "integer", "distribution": "binomial", "n": 10, "p": 0.3}  # successes

Conditional & correlated shapes#

Distributions describe a single column's marginal. To make columns relate:

Parameter aliases#

A couple of names are accepted both ways so hand-written and tool-generated schemas both work:

  • Poisson rate: lambda or lam.
  • Zipf shape: a; Pareto shape: alpha.

In the studio#

The column Inspector auto-suggests a distribution from the column name (e.g. views → power-law, price → log-normal) as a one-click preset, and the Engine params panel exposes the full distribution picker with the relevant shape fields.

Last updated Edit this page.md