Open-source agentic AI for transaction monitoring. Just ask Jooby.What's Jube? →
jube

Methodology and Taxonomy

The original guide opened with a module of its own before any R procedure: a taxonomy for what kind of analytics work is being attempted, a general project methodology, and a case for approaching the work quantitatively rather than by intuition. None of that is a technique with an R script to re-run, so it has no home among the procedure-and-equivalent pages that follow. This page states it once, up front, and points at the pages where the rest of the guide already carries it out in practice.

Descriptive, predictive and prescriptive analytics

The original guide framed a data-driven organisation as a chain: data, then reporting, then analysis, then action, then value, staffed along the way by roles from data engineer through statistician to data scientist. The analysis step itself was split into six kinds — descriptive, exploratory, inferential, predictive, causal and mechanistic — of which this guide, like the original, is overwhelmingly concerned with the predictive kind, with descriptive and prescriptive as the other two poles actually built out here.

A data-driven organisation Descriptive Exploratory Inferential Predictive Causal Mechanistic Data Reporting Analysis Action Value Data Analyst Statistician Data Engineer Business Analyst Data Scientist Quants Analysis splits into six kinds of technique; the chain is staffed by six roles along the way

This guide’s own structure already follows that three-way split, page by page rather than by declaring it:

  • Descriptive — Summary Statistics (distribution, spread, shape) and Charting (visual description) cover what already happened in a dataset.
  • Predictive — every page from Linear Regression through Neural Networks, plus the twenty-two families in Other Model Types, scores what a given record is likely to do.
  • Prescriptive — Prescriptive Analytics (Monte Carlo simulation and Bayesian-network intervention) asks what should change and what follows if it does, across a simulated population rather than one record.

Quantitative reasoning over intuition

The original guide made a deliberate case for preferring what the data says over what an analyst assumes it says, naming a working list of the cognitive biases that make unaided intuition unreliable: the conjunction and disjunction fallacies, the planning fallacy, the gambler’s fallacy, availability and representativeness heuristics, anchoring and adjustment, confirmation bias, hindsight bias, optimism bias and status quo bias among them. The point of naming them is not to memorise the list, but to justify the guide’s own working discipline: a claim about a technique’s behaviour is only as good as an actually-run result. That is also why every R snippet elsewhere in this guide is a genuine, captured console output rather than a plausible-looking transcription, and why every AdvancedAnalytics equivalent names a real class or tool rather than a plausible-sounding one — the same discipline the original guide was arguing for in its own first module, applied here to how the guide itself is written.

Cognitive biases against quantitative reasoning Fallacy Conjunction Disjunction Planning Gambler's Insensitivity to sample size Heuristics Availability Representativeness Anchoring and adjustment Bias Confirmation Hindsight Optimism Status quo Durability Conservatism Intuition Quantitative reasoning is preferred because each of these runs unchecked without it

The data / abstraction / model / activation / feedback loop

The original guide described predictive analytics as a five-stage loop — data, abstraction, model, activation, feedback — and that loop is not just a historical framing device: it is the same loop Jube’s own rule and model pipeline runs today, invocation order and all.

The data / abstraction / model / activation / feedback loop Data Abstraction Model Activation Feedback a labelled outcome feeds back into the next pass over the data
Stage What it means Where it lives in this guide and the product
Data The raw records available to reason over Loading, Shaping and Merging Data, Data Frames vs the Packed Array
Abstraction Deriving variables from raw data before modelling Abstraction; Jube’s own Abstraction Rules and Abstraction Calculations
Model Fitting a regression, classifier, clusterer or anomaly detector Every page from Linear Regression onward, and Other Model Types
Activation Turning a model’s output into a decision Activation Rules, downstream of Abstraction in the fixed invocation order AbstractionHarmonised.md describes
Feedback Recording the real outcome so it can be learned from again A labelled disposition recorded against a case, which is where a training dependent variable actually comes from (see AbstractionHarmonised.md Procedure 1)

Reading any one page in this guide against this loop places it: most pages sit in “Model,” a handful sit in “Data” or “Abstraction,” and Prescriptive Analytics is a deliberate detour off the loop — simulating many hypothetical passes through Model rather than running one real record through it.

Limits of predictive analytics

The original guide was equally direct about what a fitted model cannot do: it cannot understand a problem domain the way a person with domain experience can, and it cannot critically think or innovate beyond the data it was shown. Its stated best use is to initiate human action — surfacing a case, a score or a ranked list for a person to act on — not to replace the judgement behind that action entirely. That is the same reasoning behind Abstraction’s own claim that feature engineering, not model fitting, is where most of a predictive analytics project’s real effort belongs: a model can only ever be as good as the domain understanding encoded into the variables it is given.

Types of predictive analytics

The original guide grouped every technique it taught into a small number of shapes, and this guide’s individual pages are that same taxonomy in practice rather than a new one:

Types of predictive analytics Classification Numeric prediction Logistic regression C5.0 decision tree Linear regression Regression tree Neural network Exhaustive (Jube) Bayesian Clustering and anomaly detection Neural networks and Jube's Exhaustive engine act as universal methods, standing in for either
Shape Covered by
Numeric prediction (regression) Linear Regression, the regression tree in Decision Trees and Boosting, the regressor families in Other Model Types
Classification Logistic Regression, the C5.0 tree in Decision Trees and Boosting, the classifier families in Other Model Types
Bayesian Naive Bayes, Bayesian Analysis with bnlearn
Neural network Neural Networks, described by the original guide as a universal method able to stand in for regression or classification
Clustering and anomaly detection The clustering and anomaly-detection families in Other Model Types

Problem framing and elicitation

Before any of the above, the original guide’s first practical exercise was elicitation: given a business question (its own example was telecom customer churn), write down an exhaustive list of the reasons the outcome might occur, without yet worrying about whether a column for each reason actually exists. That list is the raw material Abstraction turns into Abstraction Rules and Calculations — elicitation supplies the candidate variables; abstraction is the mechanism that computes them. Skipping elicitation and jumping straight to whatever columns are already easiest to query is the most common way a predictive analytics project quietly narrows its own ceiling before modelling has even started.