CommerceLens
Read this before citing any number

Limitations

CommerceLens is a demonstration of analytical method, not a market research report. This page lists the boundaries of what the dataset and the analyses on this site can and cannot support.

The dataset is synthetic
Every row in CommerceLens was generated by a Python script (data/generate_dataset.py) that intentionally encodes realistic relationships (e.g. add-to-cart raising purchase probability). It is useful for demonstrating analytical method end-to-end, but it was designed to show these patterns — it does not prove that a real marketplace behaves this way. Real transactional data would need to be substituted to draw operational conclusions.
The data is observational, not experimental
No variable here was randomly assigned by an A/B test. Discount exposure, campaign exposure, device, and channel are all outcomes of the generative model’s own logic (and, in a real business, of customer choice and targeting rules), not of a controlled experiment. Associations reported throughout the site describe patterns in this sample — they do not isolate a causal effect.
Correlation is not causation
Every association page on this site (Association, Decision Lab, Probability Engine) is written to avoid causal language. A statement like “purchase probability is higher among discount-exposed sessions” describes a conditional relationship in the data; it does not mean the discount caused the purchase. Confounding variables (e.g. a customer’s underlying intent driving both discount-seeking and purchase) are plausible and were not ruled out.
Empirical probabilities depend on the sample
Every probability shown (P(Purchase | Cart), independence checks, etc.) is computed from this specific 60,000-session sample. A different sample — or the same population at a different point in time — would produce different empirical probabilities. None of these figures should be read as a fixed population parameter.
Independence checks are empirical comparisons, not proofs
The Independence Check page compares P(A∩B) to P(A)·P(B) in the observed sample. A small difference is evidence consistent with independence under this sample — it is not a formal statistical test (e.g. a chi-squared test with a p-value) and should not be presented as definitive proof that two events are independent in general.
Binomial assumptions may not perfectly hold in real e-commerce traffic
The Conversion Forecast Simulator assumes a fixed number of trials, exactly two outcomes, a constant success probability, and independent/IID trials. Real traffic violates several of these: conversion probability shifts with seasonality and campaigns, and a single customer’s repeat sessions are not independent of each other. The simulator is a useful planning approximation, not an exact model of live traffic.
The exponential assumption is a modeling choice, not a verified fact
The Continuous Distribution Lab models time-between-purchases as exponential for illustration, including its memoryless property. Real purchase intervals often show clustering, trend, and habit effects that a plain exponential model does not capture. The lab compares the model’s mean to one empirical anchor (observed average days since last order) but does not claim a formal goodness-of-fit test was run.
The triangular distribution is assumption-driven
The delivery-time triangular model is parameterized by a chosen minimum, most-likely, and maximum estimate — a common approach when historical data is sparse — rather than fit to a large empirical delivery-time sample. Its shape is only as good as those three inputs.
Distributional and independence assumptions throughout
Several calculations (binomial, hypergeometric, uniform, exponential) rest on textbook assumptions — fixed probability, sampling without replacement from a known finite pool, uniform density over a bounded interval, or independent trials. Each of these is stated explicitly on its page and flagged as a Model Simulation where it is a hypothetical forward-looking scenario rather than a re-derivation from raw rows.
Synthetic relationships were generated, not discovered
Because the relationships in this dataset (e.g. discount exposure raising purchase probability) were written into the generator, finding them again through analysis demonstrates correct method, not discovery of a real marketplace insight. The purpose of CommerceLens is to showcase the analytical workflow — descriptive statistics, association, probability, distributions, decision framing — end to end on a dataset with known, intentional structure.
In one sentence: every number on this site is arithmetically correct given the dataset it was computed from, and every dataset relationship was intentionally generated to be realistic — but “realistic” is not the same as “real,” and no claim here should be treated as evidence about an actual marketplace, actual customers, or actual causal effects.