Skip to content

Report design

The tearsheet and study board put estimates, uncertainty and sample coverage together. This page explains their choice and arrangement of information.

Summary metric selection

The tearsheet shows every configured horizon so readers can compare them without a horizon being selected for its performance.

The overview uses rank IC and the difference between the highest and lowest score buckets. Detailed panels show Pearson, cosine, variation across assets and changes over time. A missing statistic stays missing; sample and inference statuses explain why.

The CFA Institute's active management framework uses IC to describe predictive skill. Alphalens reports mean IC and differences between quantile mean returns. Together these motivate the summary's ordering and return-magnitude measures. The exact compact selection is a report design choice, not a universal KPI standard.

Temporal stability remains a detailed diagnostic; it could join the overview once its aggregation across assets and time is clearly defined. Portfolio performance measures fall outside the scope of the evaluator.

The report omits p-values. A study board shows them: the cross-sectional, time-series and cosine groups each give a p-value beside its adjustment for the recorded trial count. The underlying result tables retain the inference statistics.

Comparing trials

The list of trials is read like a leaderboard and borrows from public ones, such as Arena's agent leaderboard and BenchLM: a row of summary tiles above the table, a rank in the first column, the horizon as a segmented control, and one graphic in each row that carries the headline number. Here the graphic is the rank IC with its 95% interval on a single axis, the same for every row. A large but imprecise estimate then looks different from a small, precise one, and an interval that crosses zero is seen before any number is read.

Two choices go against what a leaderboard usually does. The tile for the highest rank IC always shows its adjusted p-value, because the best of many trials is picked by being the best. And rows are ranked only when the list is sorted by a statistic, so the newest-first order is never mistaken for merit. The study-adjustment explanation defines which displayed quantities are adjusted.

Interface design

Essential results stay visible. Optional diagnostics use one level of expandable panels, and horizon tabs keep long time-series sections manageable. Both pages, the tearsheet and the list of trials, follow the system's light or dark setting, and the charts take their colours from the same stylesheet variables, so they change with the page. On paper they are always light. The software architecture explains offline delivery.

The report follows the GOV.UK accordion guidance by keeping essential results visible and avoiding nested disclosures. Semantic headings and table markup follow W3C guidance on headings and table captions. Narrow-screen layouts contain horizontal scrolling to data regions, following the W3C reflow guidance. Horizon tabs follow the WAI-ARIA tabs pattern.

Presenting uncertainty

The report pairs estimates with uncertainty and coverage, following the ASA statement's emphasis on context. It does not label a signal as profitable because an interval excludes zero. This presentation is a design choice; it does not imply that the ASA endorses this report layout. The inference page explains what an interval means.

References