<div class="section-label"><span class="num">05</span> Relevance Metrics</div> <div class="section-title">Score <em>relevance </em>without predicting.</div> Sometimes you want the building blocks on their own: how similar each past case is to your circumstances, how informative those cases are, and how they combine into relevance, without forming a full forecast. These scores are what Predict, MaxFit, and Grid use internally when they decide which observations enter the weighted average. <p>&nbsp;</p> ## When to use Relevance Metrics - You want relevance, similarity, or informativeness as analysis outputs - You already have (or can omit) an inverse covariance and only need the score vectors - You want several scores filled in one pass > [!note] > Need a full forecast? Prefer [[Grid Prediction]] for primary analysis. Use [[Predict]] or [[MaxFit]] when you need those building-block paths. The same scores can also appear on prediction results. See [[Results/Weights & Insights|Weights & Insights]]. ## What you call ```text relevance(X, theta, cov_inv?) → scores length N similarity(X, theta, cov_inv?) → scores length N info_x(X, cov_inv?) → scores length N # no theta info_theta(X, theta, cov_inv?) → score length 1 relevance_metrics(X, theta, cov_inv?) → optional: relevance, similarity, info_x, info_theta ``` <span class="mono">X</span> is N × K. <span class="mono">theta</span> has length K (except for <span class="mono">info_x</span>). Optional <span class="mono">cov_inv</span> is K × K; if you omit it, the engine estimates it from <span class="mono">X</span>. With <span class="mono">relevance_metrics</span>, you can request only the scores you need. ## What the scores mean | Score | Shape | Idea | | --- | --- | --- | | Similarity | N | How close each training row is to <span class="mono">theta</span> | | Informativeness of X (<span class="mono">info_x</span>) | N | How unusual each row is vs the sample mean | | Informativeness of θ (<span class="mono">info_theta</span>) | 1 | How unusual <span class="mono">theta</span> is vs the sample mean | | Relevance | N | Combines similarity with informativeness of the row and of <span class="mono">theta</span> | When you run Predict / MaxFit / Grid, censoring chooses **relevance** or **similarity** (or both on MaxFit/Grid) as the score that decides include vs exclude. <div class="btn-row center"> <a class="btn-primary" href="/Functions/Grid%20Prediction">Prefer Grid Prediction</a> <a class="btn-ghost" href="/Config/Options#PredictOptions">PredictOptions</a> </div> ## Related - [[Predict]]: uses these scores inside the forecast - [[Config/Options#PredictOptions|PredictOptions]]: <span class="mono">censor_type</span> picks which score drives masking - [[Results/Weights & Insights|Weights & Insights]]: scores stored on a prediction result