Automated Data Analysis
What automated analysis actually does to a dataset, where it helps, and where it quietly goes wrong.
Automated data analysis covers everything a system does between receiving a dataset and producing an interpretation: profiling the columns, choosing a method, fitting it, checking it, and describing the result. The parts that are easy to automate — counting, summarising, plotting — are largely solved. The parts that decide whether an analysis is any good are not.
What can be automated well
Structural description: types, cardinality, missingness, distributions, obvious outliers, correlations, and simple trend and seasonality indicators. A profiler will do this faster and more consistently than a person, and will not get bored on the fortieth column.
What resists automation
Framing the question. Deciding which relationships are plausible rather than merely present. Knowing which missing values are missing at random and which encode something. Recognising that a change in a metric is a change in how the metric is collected. These require knowledge that is not in the file.
The useful middle
The most valuable systems are the ones that automate the description and then hand back something a person can argue with — a stated structure, an interval, and a note of where the model and the data disagree. That is the standard the original research set, and it remains a good one.
In this section
Automated Data Analysis
What Is Automated Exploratory Data Analysis?
Automated EDA explained: what profilers measure, the failure modes, and how to use the output properly.
How Automated Statistical Analysis Works
The pipeline behind automated statistical analysis: profiling, model grammar, search, scoring, criticism and translation.
What Is an Automatic Statistician?
An automatic statistician searches a space of models, scores them, and explains the winner in language a person can check.