{% extends "base.html" %} {% macro region(content, label, kind='table-scroll') -%}
{{ content | safe }}
{%- endmacro %} {% macro method(anchor, label='Methods') -%} {{ label }} {%- endmacro %} {% macro cell(item) -%} {{ item.value }} {%- if item.detail %}{{ item.detail }}{% endif %} {%- endmacro %} {% macro metrics_table(view) -%}
{% for horizon in view.horizons %}{% endfor %} {% for group in view.groups %} {% for row in group.rows %} {% for item in row.cells %}{% endfor %} {% endfor %} {% endfor %}
{{ view.label }} by forecast horizon
Metric{{ horizon }}
{{ group.label }}{% if group.link %} · {{ method(group.link, 'definition') }}{% endif %}
{{ row.label }}{{ cell(item) }}
{%- endmacro %} {% macro disclosure(id, label) -%}
{{ label }}
{{ caller() }}
{%- endmacro %} {% block title %}{{ title }}{% if subtitle %} · {{ subtitle }}{% endif %}{% endblock %} {% block head %}{% endblock %} {% block body %}
{% if eyebrow %}

{{ eyebrow }}

{% endif %}

{{ title }}

{% if subtitle %}

{{ subtitle }}

{% endif %}

Study design

What was tested, and on which data. Every result below refers to this window, these horizons and these return targets.

Window (UTC, end exclusive)
{{ window }}
Forecast horizons
{{ config.horizons | join(', ') }} bars{% if bar %} of {{ bar }}{% endif %}
Execution delay
{{ config.delay }} bar{{ '' if config.delay == 1 else 's' }}
Score buckets
{{ config.quantiles }} quantiles
Return targets
{{ targets }} · {{ method('method-design', 'design') }}
{% for label, value in context.items() %}
{{ label }}
{{ value }}
{% endfor %}

At a glance

Mean Spearman rank IC per forecast horizon, with its 95% confidence interval, and the spread between the top and bottom score quantiles. An interval above zero means higher scores tended to precede higher returns. Intervals account for dependence but not for other signals you tried.

{% for row in glance %}{% for item in row.cells %}{% endfor %}{% endfor %}
Mean Spearman rank IC and quantile spread by forecast horizon, cross-sectional and time-series
HorizonCross-sectional · detailsTime-series · details
Mean rank ICQ{{ config.quantiles }} − Q1 ({{ cs_units }})Mean rank ICQ{{ config.quantiles }} − Q1 ({{ ts_units }})
{{ row.label }}{{ cell(item) }}
{% if invalid_horizons %}

No valid cross-sectional observations for horizons {{ invalid_horizons | join(', ') }}. See coverage and sample status.

{% endif %}

Cross-sectional analysis

Does the signal rank assets? At each timestamp, scores are compared with subsequent returns across assets. Return target: {{ cs_metrics.target }}. {{ cs_metrics.weighting }}

{% if cs_available %}

Quantile returns

{{ method('method-quantiles') }}
{% if cs_quantiles %}{{ region(cs_quantiles, 'Cross-sectional quantile returns', 'chart') }} {% else %}

No cross-sectional quantile returns are available for this evaluation.

{% endif %}
{% else %}

No valid cross-sectional observations. Cross-sectional IC needs at least {{ config.min_assets }} usable pairs with varying scores and returns at a timestamp. See coverage and sample status for exclusions.

{% endif %}

Metrics by horizon

{{ method('method-summary') }}

Each column's returns span that horizon and are not annualised.

{{ metrics_table(cs_metrics) }}

Diagnostics

{% if cs_available %} {% call disclosure('cs-monthly', 'Monthly mean IC') %} {{ method('method-stability') }} {{ region(cs_figures.monthly, 'Cross-sectional monthly mean IC', 'chart') }} {% endcall %} {% call disclosure('cs-history', 'IC over time') %}

Rolling window: {{ options.rolling_window }} · Minimum observations: {{ options.min_periods }} · {{ method('method-stability') }}

{{ region(cs_figures.timeseries, 'Cross-sectional IC time series', 'chart') }} {% endcall %} {% endif %} {% call disclosure('cs-coverage', 'Coverage over time') %} {{ method('method-coverage') }} {{ region(diagnostics.coverage, 'Cross-sectional coverage time series', 'chart wide') }} {% endcall %} {% call disclosure('cs-status', 'Sample status') %} {{ method('method-coverage') }} {{ region(diagnostics.statuses, 'Cross-sectional sample status', 'chart') }} {% endcall %}

Time-series analysis

Does the signal time each asset? Within each asset, scores are compared with its own subsequent returns. Return target: {{ ts_metrics.target }}. {{ ts_metrics.weighting }}

{% if ts_available %}

Quantile returns

{{ method('method-quantiles') }}
{% if ts_quantiles %}{{ region(ts_quantiles, 'Time-series quantile returns', 'chart') }} {% else %}

No eligible asset quantile returns are available for this evaluation.

{% endif %}
{% else %}

Time-series IC is unavailable. {% if ts_evaluated %}No asset meets the sample and signal requirements. See the asset coverage under each horizon below. {% else %}This saved evaluation has no time-series tables.{% endif %}

{% endif %} {% if ts_evaluated %}

Metrics by horizon

{{ method('method-summary') }}

Each column's returns span that horizon and are not annualised. Intervals describe the mean, not dispersion between assets.

{{ metrics_table(ts_metrics) }}
{% endif %} {% if horizons %}

Diagnostics by horizon

{{ method('method-inference') }}
{% for horizon in horizons %}{% endfor %}
{% for horizon in horizons %}
{% if ts_available %} {% call disclosure('timing-h' ~ horizon.horizon ~ '-assets', 'Asset-level IC estimates') %} {{ region(horizon.figures.assets, 'Asset Spearman IC · ' ~ horizon.label, 'chart') }} {{ region(horizon.table, 'Asset Spearman IC estimates · ' ~ horizon.label) }} {% endcall %} {% call disclosure('timing-h' ~ horizon.horizon ~ '-monthly', 'Monthly IC and eligible assets') %} {{ method('method-stability') }} {{ region(horizon.figures.monthly, 'Asset monthly Spearman IC · ' ~ horizon.label, 'chart') }} {{ region(horizon.figures.breadth, 'Monthly eligible assets · ' ~ horizon.label, 'chart') }} {% endcall %} {% call disclosure('timing-h' ~ horizon.horizon ~ '-rolling', 'Rolling IC') %}

Window: {{ ts_options.rolling_window }} · Step: {{ ts_options.rolling_step }} · {{ method('method-stability') }}

{% if horizon.rolling_assets %}
{% endif %} {{ region(horizon.figures.rolling, 'Asset rolling Spearman IC · ' ~ horizon.label, 'chart') }} {% endcall %} {% if ts_pearson %} {% call disclosure('timing-h' ~ horizon.horizon ~ '-pearson', 'Asset-level Pearson IC') %} {{ method('method-correlation') }} {{ region(horizon.pearson_table, 'Asset Pearson IC estimates · ' ~ horizon.label) }} {% endcall %} {% endif %} {% endif %} {% call disclosure('coverage-h' ~ horizon.horizon ~ '-assets', 'Asset coverage and exclusions') %} {{ method('method-coverage') }} {{ region(horizon.coverage_table, 'Asset sample coverage · ' ~ horizon.label) }} {% endcall %} {% call disclosure('coverage-h' ~ horizon.horizon ~ '-monthly', 'Monthly coverage and exclusions') %} {{ method('method-coverage') }} {{ region(horizon.monthly_table, 'Monthly sample coverage · ' ~ horizon.label) }} {% endcall %}
{% endfor %}
{% endif %} {% if cosine %} {% call disclosure('average-position', 'Cosine similarity by asset') %} {{ method('method-cosine') }} {% for horizon in cosine %}

{{ horizon.label }}

{{ region(horizon.table, 'Asset cosine similarity · ' ~ horizon.label) }} {% endfor %} {% endcall %} {% endif %}

Signal persistence

How much do the scores and their ordering change from bar to bar? These diagnostics describe the signal, independently of forward returns; they are not traded turnover or costs.

{% if persistence %}

Persistence by lag

{{ method('method-persistence') }}

Lags count bars of the price grid. Beneath each value: contributing timestamps (cross-section, buckets) or assets (time series) of all candidates.

{% for lag in persistence.lags %}{% endfor %} {% for group in persistence.groups %} {% for row in group.rows %}{% for item in row.cells %}{% endfor %}{% endfor %} {% endfor %}
Signal persistence means by lag
MeasureLag {{ lag }}
{{ group.label }}
{{ row.label }}{{ cell(item) }}
{% for name, label in [('assets', 'Signal changes by asset'), ('periods', 'Rank persistence and universe changes by timestamp'), ('quantiles', 'Bucket membership changes by timestamp')] %} {% call disclosure('persistence-' ~ name, label) %}

Up to the first 200 rows. The complete tables are in the evaluation and its saved CSV files.

{{ region(persistence.tables[name], label) }} {% endcall %} {% endfor %} {% else %}

This saved evaluation has no signal persistence tables.

{% endif %}

Methods and interpretation

Summary metrics

Every configured forecast horizon is reported separately; none is chosen by its performance. Mean Spearman IC measures predictive ordering; its 95% confidence interval describes estimation uncertainty. Cross-sectional estimates weight usable timestamps equally; time-series estimates weight eligible assets equally. The tables use stored estimates and intervals without rerunning inference.

Intervals do not adjust for the number of signals tried. A study's list of trials reports p-values next to that count.

The cross-sectional ratio of mean to standard deviation is computed over timestamp ICs and is not annualised. The time-series median and middle half (25th to 75th percentile) describe dispersion between assets, not uncertainty.

The quantile spread subtracts the mean return of Q1 from that of Q{{ config.quantiles }}, matching the endpoints of the quantile-return chart. Each quantile mean weights its occupied timestamps (cross-sectional) or eligible assets (time-series) equally. The two endpoints can have different contributing samples. A missing extreme quantile makes the spread unavailable. Raw returns are in basis points; volatility-scaled values are in units of horizon volatility (σ). Spreads are descriptive differences in forward returns, without annualisation, trading costs or a confidence interval. They are not realised portfolio returns.

Mean pair coverage measures data availability using each analysis's own return target. Usable timestamps and eligible assets give the Spearman sample next to its total candidate count. High coverage alone does not establish predictive power.

{{ method('at-a-glance', 'At a glance ↑') }} · {{ method('method-inference', 'Statistical inference') }}

Evaluation design

Scores are compared with subsequent returns. Higher scores are expected to predict higher returns. Cross-sectional analysis compares assets at each timestamp; time-series analysis compares observations within each asset.

A bar is one row of the price data. The execution delay precedes the forecast horizon, which specifies the number of bars in the return. The evaluation window uses UTC timestamps and an exclusive end. Volatility-scaled returns express the move relative to estimated asset volatility.

{{ method('top', 'Study design ↑') }}

Information coefficient

Spearman IC
Rank correlation between scores and subsequent returns, from −1 to +1. Positive values indicate that higher scores tended to precede higher returns; zero indicates no rank association, and negative values indicate reversed ordering. IC is neither a percentage return nor a trading win rate.
Pearson IC
Linear correlation between score magnitudes and returns. Outliers and nonlinear relationships can make it differ from Spearman IC.
Mean IC
Cross-sectional estimates weight usable timestamps equally. Time-series estimates first calculate IC within each eligible asset, then weight those assets equally.
Positive timestamps or assets
The fraction of usable timestamps (cross-sectional) or eligible assets (time-series) with positive IC. This describes the prevalence of positive association, not the fraction of profitable trades.
{{ method('cross-section', 'Cross-sectional results ↑') }} · {{ method('time-series', 'Time-series results ↑') }}

Statistical inference

The 95% confidence interval describes uncertainty in the estimate. An interval spanning zero leaves the direction uncertain; narrower intervals indicate greater precision. Tables list each interval as its stored lower and upper bounds, including asymmetric intervals. An unavailable interval appears as “Unavailable”, and its inference status says why.

Cross-sectional intervals use Newey–West estimates to account for dependence, including overlapping returns. Time-series intervals use block bootstrap samples, preserving runs of paired observations and sampling assets together. The summary interval describes uncertainty in the mean, not dispersion between assets. Asset-level plots show point estimates and their individual intervals.

These intervals do not account for selecting a signal after trying many candidates.

Quantile estimation

Scores are assigned to {{ config.quantiles }} quantiles, ordered from lowest to highest. An increasing mean-return profile supports the intended ordering. Ties stay together; empty quantiles stay empty. Cross-sectional quantiles are formed separately at each timestamp, then timestamps receive equal weight within each occupied quantile.

Time-series quantiles are formed within each eligible asset's full sample, then assets receive equal weight within each occupied quantile. These retrospective groups are not thresholds that could have been traded. Return targets exclude trading costs.

Temporal stability

Cross-sectional monthly cells average timestamp-level ICs within each UTC month. Hover for usable sample counts. The rolling mean spans {{ options.rolling_window }} and requires {{ options.min_periods }} usable observations. Missing observations remain gaps, including in the rolling mean.

{% if ts_evaluated %}

Time-series monthly IC is recomputed within each complete UTC month. Both the score and return endpoint must lie inside the month. The eligible asset count may change each month. This differs from averaging timestamp-level cross-sectional ICs.

Rolling time-series IC is recomputed over trailing {{ ts_options.rolling_window }} windows, displayed every {{ ts_options.rolling_step }}. Only complete windows are used; returns must end before the displayed endpoint. Choose an asset in the menu above the chart. Empty cells are unavailable.

{% endif %}
{% if ts_evaluated %}

Cosine similarity

If a score represents a position, cosine similarity measures its uncentred alignment with returns. The timing component reflects changes in the position; the static component reflects the average position. A large static component can explain alignment without successful timing.

Cosine retains the average score and return instead of subtracting them. It measures a different association from rank correlation. These are alignment values, not portfolio returns. A constant nonzero position can have a valid cosine. Long fraction is the share of usable observations with a positive position.

{{ method('ts-metrics', 'Time-series metrics ↑') }}
{% endif %}

Sample coverage and eligibility

Cross-sectional pair coverage is usable score/return pairs divided by assets with a price, averaged equally across timestamps. At least {{ config.min_assets }} valid pairs and nonconstant scores and targets are required for a valid timestamp.

Coverage charts distinguish available assets (with prices), scored assets (with signals), and paired assets (with usable scores and returns). Each timestamp contributes one sample status per horizon.

{% if ts_evaluated %}

Time-series pair coverage divides each asset's usable pairs by its expected observations, then averages across all assets, including ineligible ones. It uses the actual time-series target, including volatility warmup losses for scaled returns.

Time-series eligibility requires at least {{ ts_options.min_pairs }} usable pairs, {{ '%.0f' | format(ts_options.min_coverage * 100) }}% coverage, and a sample spanning {{ ts_options.min_horizons }} return horizons. Monthly and rolling samples must meet these rules within their own windows.

Coverage and exclusion tables use the Spearman sample. Other methods can have different eligibility, recorded in their asset tables.

{% endif %}

Sample requirements do not establish predictive power. Sample status explains missing correlations; inference status explains missing intervals. A dash denotes an unavailable statistic, never zero. Final bars may lose return labels when their horizons reach the evaluation boundary.

{{ method('cs-coverage', 'Cross-sectional coverage ↑') }}

Signal persistence

Diagnostics use scores and prices available at each endpoint, independently of forward-return horizons. Lag is measured in price-grid bars; a multi-bar lag compares endpoints and does not sum intervening changes.

Rank persistence is the cross-sectional rank correlation of scores with their values one lag earlier, from −1 to 1, weighting valid timestamps equally. Bucket entry fraction is the share of a bucket's current members absent from it one lag earlier; ties stay together. Universe counts reflect availability of both a score and a price, not confirmed listings or delistings.

Normalised change divides each asset's total absolute score changes by its total mean endpoint magnitudes on the same valid pairs, then weights eligible assets equally. A constant nonzero signal has zero change; an all-zero signal has undefined normalised change. Sign changes exclude transitions into or out of zero, which are counted separately.

Missing values remain unavailable; a gap never shortens a lag. No confidence intervals are claimed.

{{ method('signal-persistence', 'Signal persistence ↑') }}
{% endblock %}