top of page

THE METHODOLOGY BEHIND CSA

Relevance-Based Prediction

Transparent prediction, task by task. 

Draw on the historical observations that matter for each prediction. Understand the evidence behind the result and how strong it supports your decision. 

+ MODEL FREE

+ TRANSPARENT

+ TASK-SPECIFIC

INSIDE A PREDICTION                      ILLUSTRATIVE EXAMPLE

Predicted performance

Evidence supporting the result

0.58

High Reliability

Expected goals per 90

Which variables influence this prediction?

Goals per 90

Pass completion
 
Duel win rate

Other variables

Conceptual display. Values and bars illustrate the page design.

01   A DIFFERENT APPROACH

A model's performance only tells part of the story.

Simplicity can limit flexibility. Complexity can limit transparency.

SIMPLE MODELS

COMPLEX MODELS

Clear, but sometimes restrictive

Flexible, but harder to inspect

Simple models are easy to interpret, but a fixed relationship may miss differences across players and situations. 

Complex models can capture richer patterns, but their internal workings can be difficult to explain and may require substantial data and parameter estimation. 

AVERAGE MODEL QUALITY

Task-Specific Reliability

The hidden cost of averages

Models are often evaluated by their performance across many predictions. But strong overall performance does not reveal how much confidence to place in an individual prediction. For a team making a decision about a particular player, this distinction matters.

THE FOUNDATION

With RBP, each prediction has its own evidence. 

RBP forms predictions as weighted averages of historical outcomes. The weights reflect the relevance of past observations to the circumstances being evaluated. 

Prediction = Σ relevance weights × observed outcomes

Which past observations and which variables matter most for this particular prediction?

02   RELEVANCE

What makes an observation relevant?

Relevance combines two ideas to identify useful historical evidence for the circumstances at hand.

Similarity 

Informativeness

Relevance

How closely does a historical observation resemble the circumstances being evaluated?

How much does the observation differ from the average?

Combines similarity and informativeness to determine how historical observations contribute to a particular predictions. 

03   RELIABILITY

How much support does
this prediction have?

Fit measure how strongly relevance aligns with observed outcomes for the chosen task. It can be assessed before the unknown outcome occurs.

R-squared summarizes how well a regression explains outcomes across the sample. Fit brings the focus to the evidence for an individual prediction. 

Strong alignment

Strong alignment between relevance and outcomes indicates useful patterns that support a prediction.

Mixed alignment

Mixed alignment between relevance and outcomes indicates weaker patterns that support a prediction. 

Weak alignment

Weak alignment between relevance and outcomes indicates little evidence of useful patterns to support a prediction. 

Conceptual description, not numeric Fit categories. 

04   GRID PREDICTION

MANY COMBINATIONS. ONE PREDICTION.

VARIABLE COMBINATION x RELEVANCE THRESHOLDS

Fit-weighted blending

Illustrative grid. Each cell has a prediction and Fit score. 

Evaluate different combinations of information.

Which combinations of variables and observations are most informative depends on the specific prediction task. 

Grid prediction evaluates different combinations of predictive variables and relevance thresholds, each yielding its own prediction. RBP combines these predictions, giving greater weight to those with stronger Fit for the specific task.

Preserve useful information

When data is missing, RBP retains useful information from incomplete observations and accounts for what is missing when forming each prediction and assessing its reliability.

05   UNDERSTANDING EACH PREDICTION

See what drives
the result.

Explore the historical observations and variable contributions that shape each prediction, along with the evidence supporting its reliability.

/01

What is the prediction?

The prediction is a weighted average of historical outcomes, with weights determined by their relevance to the specific task.

PREDICTED OUTCOME

/02

How strong is the evidence?

Fit measures the strength of useful patterns in the data for the specific prediction, providing an advance assessment of its reliability.

TASK-SPECIFIC FIT

/03

Which observations matter?

Relevance weights show which historical observations shape the prediction and how much each contributes.

RELEVANT OBSERVATIONS

/04

Which variables influence it?

​Impact on Prediction (IOP) and Impact on Fit (IOF) reveal how each variable affects the predicted outcome and the strength of evidence supporting it.

VARIABLE IMPORTANCE

Two ways to understand a variable's contribution. 

RBP compares grid results with and without each variable to assess its effect on the predicted outcome and its reliability.

IOP / IMPACT ON PREDICTION

IOF / IMPACT ON FIT

Does it raise or lower the prediction?

Does it strengthen or weaken the evidence?

IOP measures how much including a variable raises or lowers the predicted outcome.

IOF measures how much including a variable strengthens or weakens the evidence supporting the prediction.

06   PAPERS, APPLICATIONS, & DEEP DIVES

Explore the research.

Explore the research behind Relevance-Based Prediction, applications across sports, and short explanations of the concepts behind the methodology. 

Filter by Topics

Deep Dive

Deep Dive: RBP vs Linear Regression

How RBP compares with regression

Deep Dive

Deep Dive: Informativeness

Why unusual observations matter

Deep Dive

Deep Dive: Asymmetry

Detecting nonlinear patterns in data

Deep Dive

Deep Dive: Grid Prediction

Combining predictions based on reliability

Research Paper

The Virtue of Transparency in Sports Analytics: An Application to NFL Quarterbacks

An RBP application to the NFL

Research Paper

Prediction with Conviction: An Application of Relevance-based Prediction to the NBA

An RBP application to the NBA

Deep Dive

Deep Dive: Handling Missing Data

How RBP handles missing data.

Research Paper

Prediction with Incomplete Information

How RBP handles incomplete information.

Deep Dive

Deep Dive: The Distribution of Information

Confidence when past examples disagree

Deep Dive

Deep Dive: Similarity and Relevance

How RBP selects historical observations

Deep Dive

Deep Dive: Relevance-Based Importance

Which variables improve prediction reliability?

Deep Dive

Deep Dive: Fit

How reliable is a prediction?

Research Paper

Prediction with Transparency – Offensive Value in Baseball

An RBP application to the MLB

Research Paper

The Virtue of Transparency: How to Maximize the Utility of Data Without Overfitting

How RBP handles overfitting

Research Paper

A Transparent Alternative to Neural Networks with an Application to Predicting Volatility

A comparison of relevance-based prediction and neural network approaches.

Deep Dive

Deep Dive: Predicting Binary Outcomes

How RBP predicts binary outcomes.

07   APPLYING RBP

Explore RBP with your own data.
 


Use the Prediction Lab's no-code interface, incorporate the Prediction Engine into your analysis, or work with CSA on a question specific to your organization. 

©2026 Cambridge Sports Analytics, a division of Cambridge Prediction Analytics, LLC. All rights reserved.      

Cambridge, Massachusetts

bottom of page