Research Analysis: Financial Metric Evaluation with IC, Quintile, & Threshold Analysis¶
Last updated 2026-11-07
This document describes the quantitative research methodology used in the QuantAscent quant engine for evaluating factor predictiveness, constructing composite scores, and backtesting strategies.
Who This Document Is For¶
You don't need a statistics degree to use QuantAscent's research tools — the app handles all the math for you. This document explains what the numbers mean so you can make better decisions about which metrics to include in your strategies. If you're new to quantitative investing, read the plain-language summaries (in blockquotes) at the top of each section and skip the formulas. If you want the full technical detail, it's all here too.
Using the app? See the Screen Guide for a walkthrough of the Research page and every other screen.
Table of Contents¶
- Overview & Motivation
- Data Preparation & Forward Returns
- Information Coefficient (IC) Analysis
- Quintile Analysis
- Rule Preview
- Testing a Filter
- Where It Works: IC by Group
- Composite Scoring Systems
- Backtest Performance Metrics
- Metric Computation Reference
1. Overview & Motivation¶
The research framework answers a fundamental question in quantitative investing:
Given a universe of stocks and a set of observable metrics, which metrics reliably predict future returns, and how should we combine them into a selection strategy?
The pipeline proceeds in stages:
Raw Data Preparation --> Financial Metric Calculation --> Metric Evaluation -->
Composite Score Generation --> Backtest --> Strategy Optimization
Each stage introduces statistical tools designed to separate genuine predictive signal from noise. The key analyses — IC ranking, quintile decomposition, the rule preview, the filter funnel and IC by group — each attack this problem from a different angle, providing converging evidence about metric quality.
2. Data Preparation & Forward Returns¶
In plain language: Before we can test which metrics predict stock returns, we need to organize the data. The "metrics matrix" is a big spreadsheet: each row is one stock on one date, and the columns are 150+ financial measurements (like P/E ratio, ROE, debt levels, etc.) plus the stock's actual return over the next quarter. This is the raw material every analysis in this document builds on.
2.1 The Metrics Matrix¶
The foundation of all analysis is a cross-sectional panel (the "metrics matrix"): a tabular dataset where each row represents one (ticker, as_of_date) observation, and columns contain 150+ fundamental, valuation, growth, and technical metrics plus a target variable.
Construction:
- Rebalance dates are generated at fixed intervals (default 90 days) across the historical sample. The Build Matrix panel defaults the start date to January 2016 — the beginning of the window where the database's delisted-company coverage is complete (see the survivorship note in Section 9.1). You can build over an earlier window; the app shows a warning explaining that those results are survivorship-biased.
- For each rebalance date, metrics are computed for every ticker with sufficient data as of that date.
- The result is augmented with categorical company info (sector, industry, country, exchange).
2.2 Forward Return (Target Variable)¶
The dependent variable in all analyses is FutureProfit — the simple holding-period return over the rebalance window:
where:
- \(P_t\) = closing price on the rebalance (observation) date
- \(P_{t+\Delta}\) = closing price on the date \(\Delta\) days forward (default \(\Delta = 90\))
- If no price exists on exactly \(t + \Delta\), the most recent available price on or before that date is used.
Outlier handling: FutureProfit values are winsorized — clipped, not dropped — to a horizon-scaled cap of \(\pm(10.0 \times \Delta/365 + 1.0)\), about \(\pm 345\%\) for the default 90-day window. This prevents extreme values from dominating means while keeping real right-tail winners and \(-100\%\) bankruptcies in the sample, so research statistics are measured on the same sample the backtester trades.
2.3 Growth Rate Estimation¶
Many metrics involve estimating a growth rate from a time series of historical observations (e.g., FCF over the last 20 quarters). This is done via OLS regression on normalized values:
Given a series of \(n\) observations \(y_1, y_2, \ldots, y_n\) (ordered newest-first from the data source):
Fit the linear model \(\tilde{y}_i = \beta_0 + \beta_1 \cdot i + \varepsilon_i\) via OLS:
The goodness-of-fit is captured by \(R^2\):
A high \(R^2\) indicates the growth trend is consistent (not noisy), which is itself a useful signal.
2.4 Describing a Universe¶
Plain language: before asking which metrics work inside a universe, the Universe step says what the universe is. A strategy's result is what the whole universe did plus what the picking added, so the two are measured separately and never folded into one score.
All of these are computed on the metrics matrix, per rebalance date, inside the universe query (the same query the Backtester's Filtered Universe line uses):
- Breadth — companies per date. Reported as the median quarter and the thinnest quarter.
- Base rate — the per-date equal-weight mean forward return, compounded at \(365.25 / \text{rebalance days}\) periods a year (not a flat 4), with annualized volatility and the maximum peak-to-trough drawdown of that path. Every company in the universe on a date is held, however few, and a date on which the universe is empty is a 0% period (carried flat, as the Backtester's Filtered Universe line does). The whole-market reference is the same statistic on every company passing the freshness screen, aligned on the universe's dates; the SPY reference is SPY's dividend-adjusted return over the same forward windows.
- Single-pick odds — per company-quarter: the share with a positive forward return, the mean gain when positive, the mean loss when negative, the share at or above +100%, the share at or below −50%.
- Dispersion — per date, the 80th minus the 20th percentile of forward return, averaged across dates. The opportunity a ranking can harvest.
- Stable signals — the date range is split in half; a metric is stable when its per-date IC (§3) has the same sign in both halves and each half holds at least 500 valid pairs. The Adaptive Builder's stability gate uses the same halves.
- Coverage — the share of rows carrying a value, per metric; a metric needs 30% to be ranked (§3.3).
- Honest window — delisted companies are present in the data from 2016; earlier years hold survivors only, and a universe starting before then overstates returns.
- Dollar volume —
DollarVolume_3m, the mean of close × shares traded over the trailing 63 sessions, in $M per day. Backtests charge one flat slippage rate whatever a stock's trading volume, so thin trading and wide spreads are not modeled.
The sector × size grid applies the same measures to each of eleven sectors by five market-cap bands (Mega ≥ $200B, Large \(10B–\)200B, Mid \(2B–\)10B, Small \(300M–\)2B, Micro < $300M), keeping every other filter of the universe. A cell with fewer than ten names on a typical date is not shown.
3. Information Coefficient (IC) Analysis¶
In plain language: The IC answers a simple question: "If I rank all stocks by this metric, do the higher-ranked ones actually earn higher returns?" An IC of +0.05 means yes, there's a meaningful positive relationship. An IC near zero means no relationship. A negative IC means the relationship is inverted (lower values predict higher returns). You don't need to understand the math to use the Research page — just look for metrics with higher |IC| values.
3.1 Definition¶
The Information Coefficient (IC) measures the cross-sectional rank correlation between a metric's values and subsequent realized returns. It answers: "Do stocks ranked highly on this metric tend to have higher future returns?"
We use Spearman's rank correlation rather than Pearson's because:
- It captures monotonic (not just linear) relationships.
- It is robust to outliers and non-normal distributions.
- Financial return distributions are heavy-tailed; rank-based measures are more stable.
3.2 Calculation¶
For a given metric \(M\) and a single rebalance date, the cross-sectional Spearman correlation is computed over that date's valid observations \(\{(m_i, r_i)\}_{i=1}^N\), where \(m_i\) is the metric value and \(r_i = \text{FutureProfit}_i\):
where \(d_i = \text{rank}(m_i) - \text{rank}(r_i)\) is the difference in ranks.
The reported IC is the average of these per-date cross-sectional correlations across all rebalance dates (a date needs at least 10 valid pairs for its cross-section to count) — not one pooled correlation over the whole stock-period panel. Pooling would rank returns across time periods, so any metric with temporal drift would absorb spurious IC from regime effects, and a 4,000-stock period would outweigh a 40-stock one.
In practice, tied ranks are handled appropriately and a two-tailed \(p\)-value is provided.
Interpretation:
| IC Range | Interpretation |
|---|---|
| > +0.05 | Meaningful positive predictive power (higher metric → higher returns) |
| -0.05 to +0.05 | Weak or no predictive power |
| \< -0.05 | Meaningful inverse predictive power (higher metric → lower returns) |
In cross-sectional equity research, ICs in the \(\pm 0.03\) to \(\pm 0.10\) range are typical for useful single factors. An IC of \(0.10\) is considered quite strong.
3.3 IC Ranking Scan¶
The IC ranking scan computes IC for every numeric metric in the matrix, then sorts by \(|\text{IC}|\) to surface the most predictive factors. Filters applied:
- Minimum data coverage: at least 30% of observations must have non-null values for the metric.
- Optional pre-filter (e.g., restrict to a sector or market-cap band).
- Optional exclusion of dollar-denominated metrics (which may carry size bias).
Output columns:
metric: factor nameIC: Spearman correlation with FutureProfitabs_IC: \(|\text{IC}|\) (for ranking)ic_ci_lo/ic_ci_hi: 95% confidence interval bounds on the IC. With two or more rebalance dates these come from a \(t\)-test on the per-date IC series (sample size = number of dates — panel rows are not independent, so a pooled-N interval would overstate confidence). With a single date, the classical Fisher z-transform applies: \(\text{tanh}(\text{atanh}(r) \pm 1.96 / \sqrt{N-3})\)p_value: two-tailed significancen_valid: sample sizedirection: "higher is better" or "lower is better"
The 95% CI is shown as a half-width in the results table (e.g. ±0.04) and as error bars on the chart. At the unfiltered universe scale (~4000 stocks), each date's IC is measured precisely, so the CI width mostly reflects how stable the metric's IC has been across rebalance dates. Filters that narrow the universe (single sector, single mcap band) make each date's IC noisier and widen the CI — per-metric variation that's useful for distinguishing real signal from noise.
3.4 Comparison Modes (Whole Market vs. Peer-Relative)¶
The "Compare to" dropdown selects how each metric is ranked before correlating with future returns:
- Whole market (default): correlate raw metric values across the entire filtered universe. This is the same scoring rule the backtester uses, so IC values here correspond to the historical returns of a raw-rank strategy scored the same way.
- Sector peers: rank-transform each metric within sector to a (0, 1] percentile, then correlate the percentiles with future returns. This neutralizes cross-sector confounds — answers the attribution question "does this metric predict returns within sectors, or am I just picking up which sectors outperformed?".
- Sector + size peers: same as Sector peers, but the peer group is (sector × market-cap band) using the standard Mega / Large / Mid / Small / Micro thresholds. Strongest attribution control. Peer groups with fewer than 5 stocks are dropped (rank percentiles are unreliable for tiny groups).
Important interpretation note. Peer-relative IC and raw IC answer different questions and tend not to agree. A metric that looks predictive in Whole market may collapse to near-zero under Sector + size peers if its raw signal was driven by sector or size exposure rather than within-peer variation. Conversely, a metric with modest raw IC may strengthen under peer-relative analysis if it carries genuine within-peer signal that gets diluted across heterogeneous sectors.
Use the peer-relative modes to understand which metrics carry within-peer signal vs. which ride cross-sector trends. Use Whole market when the IC should be measured with the same scoring rule a running strategy uses, so the historical measurement is like-for-like. The peer-relative modes are research lenses, not strategy-construction tools.
3.5 Rationale¶
IC analysis is the first pass in factor evaluation. A factor with near-zero IC is unlikely to add value regardless of how it is incorporated. However, IC alone has limitations:
- It assumes a linear-in-ranks relationship (a factor could be predictive only in the tails).
- It does not reveal the magnitude of return separation.
- It is an aggregate measure and may mask regime- or sector-dependent effects.
These limitations motivate the complementary analyses below.
4. Quintile Analysis¶
In plain language: Quintile analysis splits the stocks into five equal groups on each rebalance date — from the lowest 20% to the highest 20% by a metric — then asks what holding each group as an equal-weight portfolio would have earned per year. If Group 5 (highest values) beats Group 1 (lowest values), the metric is probably worth using. If the returns go up step by step from Group 1 to Group 5, that's even better — it means the relationship holds across the whole range, not just at the extremes. The plain average of every stock in a group is not used: a few extreme winners can carry it the wrong way.
4.1 Motivation¶
While IC summarizes the overall rank relationship in a single number, quintile analysis reveals the shape of the return-metric relationship by grouping stocks into ordered buckets and examining average returns within each.
4.2 Portfolio Formation¶
On each rebalance date \(d\), the valid observations of metric \(M\) are sorted and divided into five equal-count groups (ties broken by order):
- \(Q_1\) = the lowest 20% of metric values on that date, \(Q_5\) = the highest 20%
- The cut is made within each date, never pooled across dates, so a drifting metric does not put its early years in one group and its late years in another
- A date with fewer than 20 valid observations is skipped
4.3 Each Fifth as a Held Portfolio¶
The statistic shown is what an investor holding the whole fifth, equal weight and re-formed every rebalance date, would have earned. On each date the fifth's return is the mean of its names,
and the per-date returns compound into a yearly rate, with \(P\) periods per year (\(365.25 / \text{rebalance days}\)) and \(D\) dates:
This is the same definition as the universe base rate and the Backtester's universe line, so the fifths, the base rate and the rule preview's per-date difference all measure returns the same way. One name's extreme quarter counts as one name's share of one date's return, which is what a holder of that fifth received.
The table adds, per fifth: the median name's forward return \(\tilde{r}_j\) (the typical outcome), the share of names with a positive forward return, the worst single period \(\min_d \bar{r}_{j,d}\), and the number of observations.
4.4 Why Not the Plain Average¶
A pooled average of every name in a fifth across every date,
weights every observation equally, so a handful of extreme winners can carry one group's average above a group whose typical name did better on most dates. In a small-cap universe this happens routinely: the lowest-margin fifth can show the highest pooled average while its median name trails every other fifth and the portfolio holding it compounds the slowest. The card shows the held return instead and prints a warning when the pooled average disagrees with it, naming the fifth and how many points its ten biggest winners add to that average.
4.5 Quality Metrics¶
Q5−Q1 spread, in points per year: \(R_5 - R_1\), the difference between holding the top fifth and holding the bottom fifth.
Order: whether \(R_1 < R_2 < R_3 < R_4 < R_5\) (or the reverse). The card states the reading in a sentence under the chart.
4.6 Interpretation Guide¶
| Pattern | Reading |
|---|---|
| Rises from Q1 to Q5 | The metric sorts the whole range: use it as a scoring criterion |
| Falls from Q1 to Q5 | Lower values are better: score it with that direction, or keep the top fifth out |
| One end stands out, the rest flat | The metric works at one end: preview a rule on that end rather than scoring the range |
| No order | Not predictive across the range; check the IC's interval before using it at all |
| The plain average disagrees with the held return | A few extreme names carry the average; trust the held return and the median |
5. Rule Preview¶
In plain language: the rule preview asks the practical question: "If I had only bought stocks where metric X was above (or below) some cutoff, what would have happened?" It applies the cutoff on every rebalance date, the way the backtester applies a filter, and reports what passed, what those stocks returned against the whole universe, and how sure that lift is. It replaced the threshold sweep, which pooled every date into one sample.
5.1 Two Kinds of Cutoff¶
Percentile of the universe keeps a share of the universe by the metric's rank: on each date \(d\) the cutoff is recomputed from that date's values,
and the mirror image for the bottom \(L\%\). Ties at the cutoff are kept. This is how a percentile floor behaves in a strategy.
Fixed value keeps \(S_d = \{i : m_{i,d} > \tau\}\) (or \(<\)) with one \(\tau\) on every date, the way a metric filter behaves.
A date whose pool has fewer than 20 valid values is skipped.
5.2 Per-Date Statistics¶
For each date: the number of names passing, the mean forward return of the passing names \(\bar{r}_{S_d}\), the mean of the whole universe \(\bar{r}_{U_d}\), and the lift
The readout reports the median names per date and the share of the universe passing in one sentence, then four tiles: the mean of the per-date pass returns, the mean of the per-date universe returns, the mean lift with a 95% \(t\)-interval on the per-date lift series, and the count of dates with a positive lift. The two-sided \(p\) of the same test is in the Difference tile's tooltip. This is the same inference the IC uses (§3.2): rows within a date share the period's shock, so dates are the independent observations, not rows. A line under the tiles reads the interval: above zero, below zero, or including it (not clearly different from the universe over these dates).
The curve over every share. The left chart repeats the percentile rule at every share kept \(s \in \{2, 5, 10, \ldots, 95, 100\}\%\) on the chosen side, in one pass per date: the valid names are sorted by the metric, the running mean of the forward return over the first \(k_s = \max(1, \operatorname{round}(s \cdot n_d))\) names is the passing return at every share at once, and the lift is that minus the date's universe mean. The chart plots the mean lift over dates against \(s\) with the 95% \(t\)-interval as a band and marks the preview's own position (its share and lift; a fixed-value rule is placed at the share it passes). Ties are broken by order here where the preview keeps them, so the two can differ by a name or two on a tied date. The right chart's y-range is the largest per-date lift over that grid on the chosen side at shares of 10% and more, so it does not change when the cutoff moves and a weaker position shows smaller bars.
5.3 Writing the Rule¶
The preview becomes a strategy rule unchanged: a percentile rule as a universe floor (percentile_floors), a fixed-value rule as a metric filter, or either as a scoring criterion (weight + direction).
6. Testing a Filter¶
In plain language: three places answer "what does this filter do?": the Universe step for the universe as a whole, the rule preview for one metric, and the filter funnel for a saved strategy's full stack.
- The universe — add the filter to the universe editor and the profile card (§2.4) says what the universe becomes: names per quarter, base rate against the whole market and SPY, single-pick odds, dispersion, stable signals.
- One metric — the rule preview (§5) applies a cutoff per date with a lift interval.
- A strategy's stack — the filter funnel on the Combine step applies a saved strategy's filters in schema order (size, price, dollar volume, freshness, categories, metric filters, percentile floors) and reports, per stage, the names left, the share left, and the equal-weight return and hit rate of what is left.
The former pooled comparison (Welch's \(t\), Mann-Whitney \(U\), Cohen's \(d\) on every row across every date) is retired for the reason §3.2 gives for the IC: pooled rows are not independent observations.
7. Where It Works: IC by Group¶
In plain language: Some metrics work well for large companies but not small ones, or in certain sectors but not others. The Signals step's Where it works card measures a metric's predictive power inside each sector and each market-cap band separately, so you can see whether a metric that looks good overall is carried by one group, and use it only where it works.
7.1 Motivation¶
A factor's predictive power may not be uniform across sectors or size buckets. IC by group computes the IC (§3) inside each group on its own:
with the same per-date inference as the overall IC (§3.2): within each group the IC is computed on every rebalance date and the dates are the observations. A group with fewer than 20 valid names on a date contributes nothing for that date. The former median-split and quintile-split variants, which conditioned one metric on the values of a second metric, were retired with the Conditional IC tool; sector and size are the two groupings the product reports.
7.2 Market Cap Band as a Grouping¶
Market cap is a continuous variable but the size-effect literature treats it as a small set of size buckets. QuantAscent exposes a virtual MarketCapBand column derived from MarketCap:
| Band | Range |
|---|---|
| Mega | > $200B |
| Large | $10B - $200B |
| Mid | $2B - $10B |
| Small | $300M - $2B |
| Micro | < $300M |
The Signals step's "Where it works" card reports a metric's IC inside each band (largest to smallest) and inside each sector, each with its own per-date inference (§3.2); the sector × size grid on the Universe step applies the same bands to the universe measures.
7.3 Score IC by Group¶
The Combine step's Score Breakdown measures the composite strategy score's IC within subgroups:
Grouping dimensions include:
- Sector/Industry — Does the strategy work uniformly across sectors?
- Country/Exchange — Geographic robustness check.
- Market Cap Band — Size-dependent effectiveness, using the bands defined in §7.2.
- Year — Temporal stability of the signal.
Minimum subgroup size of 20 observations is required to compute a meaningful IC.
8. Composite Scoring Systems¶
In plain language: Once you've identified individual metrics that predict returns, the next step is combining them into a single score. QuantAscent ranks every stock on each metric against all the others in the universe ("how does this stock rank on each metric compared to all others?"), weights the ranks and sums them. The stocks with the highest composite scores become your portfolio.
8.1 Percentile Scoring (Continuous Composite)¶
Each metric is ranked across the universe and the weighted percentile ranks are combined into a continuous composite:
where:
- \(\text{pctrank}_k\) is the cross-sectional percentile rank of metric \(k\) (0 to 100).
- Direction-aware: for "ascending" metrics, higher raw values get higher percentile ranks; for "descending," lower values rank higher.
- Optional clipping to a percentile range focuses the scoring on a relevant region of the distribution.
- \(w_k\) is the user-defined weight for metric \(k\).
The composite score ranges from 0 to 100.
8.2 Candidate Selection¶
After scoring, the candidates are sorted by composite score, descending, and the top max_stocks are taken; equal scores are ordered by ticker, so the selection is the same on every run. The hard filters (market cap range, sector exclusions, metric filters, percentile floors) apply before scoring.
8.3 Construction Mode (Peer-Relative Scoring)¶
In plain language: When percentile scoring ranks stocks, it normally compares every stock against every other stock in the filtered universe. That's a whole-market ranking. The alternative is peer-relative ranking — group stocks into peer cells first (by sector, or by sector × market-cap band), and rank each stock only against its peers before combining into a composite. Peer-relative scoring picks "the best of each peer group" rather than "the highest absolute scorers regardless of where they came from."
Scoring strategies (Section 8.1) expose a construction_mode setting with three values:
| Mode | What it does |
|---|---|
raw (default) |
Each metric is ranked across the whole filtered universe. This is how strategies behaved before this option existed. |
sector |
Each metric is ranked within sector peer groups. A Tech stock and a Healthcare stock at the same raw metric value can end up with different percentile ranks because they're being compared against different peer distributions. |
sector_size |
Each metric is ranked within (sector × market-cap band) peer groups. Mega-cap Tech compares only to other mega-cap Tech, mid-cap Healthcare only to other mid-cap Healthcare, etc. |
The market-cap band buckets match those used by IC analysis (Section 3.4): Mega ≥ $200B, Large \(10B–\)200B, Mid \(2B–\)10B, Small \(300M–\)2B, Micro < $300M. Peer groups smaller than the minimum required by the rank are still kept — percentile ranking degrades gracefully in tiny groups but does not crash.
Why this exists as an option, not a default
Empirical testing (a 60-strategy factor research set × 65 quarters of out-of-sample data) showed that peer-relative construction is approximately a wash with whole-market construction on aggregate — the mean change in Sharpe is small and unsigned. The two modes systematically differ by sector, however: within-firm-dispersion sectors (Healthcare, Communication Services, Financial Services, Utilities) tend to benefit from peer-relative scoring, while move-as-a-unit sectors (Basic Materials, Real Estate, parts of Energy) tend to be hurt by it.
Because whole-market ranking is what the validation evidence covers, the construction-mode setting is exposed as a user opt-in on a per-strategy basis. Users investigating peer-relative effects via Research → IC Rankings (Compare to: Sector / Sector + size peers) can carry that lens forward into the strategy they build, by setting the matching construction mode on the deployed strategy.
Cross-metric criteria (operator a>b or a<b, Section 8.1) are unaffected by construction mode — they score 100 or 0 on the raw values, regardless of which peer-group ranking is applied to single-metric criteria.
8.4 Redundancy Among Scoring Metrics¶
In plain language: a composite score that reads five metrics which all rank companies the same way counts one signal five times, not five. The Combine step's redundancy grid shows which pairs of a strategy's scoring metrics rank the companies alike.
For every pair of metrics the score reads (both sides of a cross-metric criterion included), the Spearman rank correlation is computed across the companies on each rebalance date, inside the strategy's filtered universe, and averaged over the dates with at least 10 rows. Each metric is first turned the way the score reads it: a metric scored lower-is-better (a criterion with direction asc, the right side of an a>b criterion) has its sign flipped, so that \(s_a s_b \rho_d(a, b)\) with \(s \in \{+1, -1\}\) is the correlation between the two metrics as the score reads them:
Per date rather than pooled, for the reason §3.2 gives for the IC: a correlation pooled across dates absorbs whatever the two metrics do together over time. A pair ranks the companies alike when \(\bar{\rho}_{ab} \ge 0.7\): the score counts one signal twice. A pair at \(\bar{\rho}_{ab} \le -0.7\) ranks in opposite orders: the score rewards what the other penalizes, and the two terms largely offset. Turning each metric first is what makes the sign readable. Without it, a low-volatility metric (lower is better) and a return-on-capital metric (higher is better) that pick the same companies show a raw correlation near −0.4 and read as diversifying when they are largely one signal. The grid describes the score's inputs and is not a verdict: two metrics that rank alike can be kept on purpose, as a deliberate overweight. Pointing at a cell also shows the pair's pick overlap (§8.5).
8.5 Independent Signals Among the Reliable Metrics¶
In plain language: the metrics at the top of the IC table are not as many ideas as there are rows. Four volatility measures that rank the same companies the same way are one idea with four names. The Signals step's Independent signals card lists one row per idea, the strongest metric first, with the metrics that rank like it beside it.
The card starts from the metrics that pass both of the IC table's reliability checks: the p-value of the per-date IC series (§3.2) is under 0.05, so the 95% range excludes zero, and the Stable mark is set (the same IC sign in both halves of the date range with at least 500 pairs in each). It takes the 20 strongest of them by \(|IC|\), inside the universe the table was computed on (the Universe step's query). Each is turned by the sign of its IC, so \(s_a = \operatorname{sign}(IC_a)\) and the oriented correlation \(s_a s_b \bar{\rho}_{ab}\) of §8.4 is the correlation between the two as the IC reads them. Then, from the strongest down:
- The first metric starts signal 1.
- Each next metric is compared with the first metric of every signal so far. If its oriented correlation with all of them is under 0.7, it starts a new signal. Otherwise it is listed under the signal whose first metric it correlates with most, and the card shows that correlation beside it.
Because a metric is compared with first metrics only, the signals' first metrics are mutually under 0.7, which is what the headline count claims. The cost is that a listed metric can be at or above 0.7 with another signal's first metric (a cross match); the card counts these under the table and names them in the metric's tooltip rather than merging the two signals. Comparing with every listed metric instead would merge through such pairs: on the October 2026 Builder matrix that puts PriceTo52wHigh, a momentum metric at 0.60 with IdioVol_1y, inside the volatility signal because it reads 0.71 with DownsideVol_60d. Two signals whose first metrics are at or below −0.7 rank in opposite orders; both stay and the pair is named.
Effective signals. The 0.7 rule is a duplicate screen: it merges a metric with another that ranks the companies the same way, and it leaves alone a set of metrics that are alike without reaching 0.7 (five profitability metrics at 0.4 to 0.7 with each other are five signals by the rule). So the headline carries a second number with no threshold in it. For the \(k\) signals' oriented correlation matrix \(R\) (the grid), the effective count is
the participation ratio of the eigenvalues of \(R\) (the trace of a correlation matrix is \(k\)). Unrelated signals count in full (\(R = I\) gives \(k\)); copies or opposites count once (\(R_{ab} = \pm 1\) gives 1); \(k\) signals all at \(\rho\) give \(k / (1 + \rho^2 (k - 1))\), so 11 at 0.4 give 4.2. The headline rounds it ("together about 4 independent signals' worth of ranking information"); the rule line under it gives one decimal. On the October 2026 Builder matrix the 11 signals have 55 pairs all under 0.7 with a mean of 0.40, and \(k_{\text{eff}} = 4.0\).
Pick overlap. The grid's hover line and the tooltips add a plain-words companion to the correlation: on each rebalance date, each metric's top tenth of the companies with a value (after the turn) is taken, and the pair's overlap is the size of the intersection over the smaller of the two sets, averaged over dates. On the October 2026 Builder matrix a correlation of 0.7 to 0.8 went with about half of the same top-tenth picks, 0.8 to 0.9 with about 62%, and 0.9 and over with about 82%, while pairs at 0.3 to 0.5 shared about 14%. The overlap is shown, never used as a rule.
The result is a description of overlap among the reliable metrics in this universe, not a recommendation. The set changes with the universe; the 0.7 threshold is the same convention §8.4 uses; the count of 20 is a display choice (on the same matrix 86 of 133 metrics passed both checks, with ICs from 0.120 down to 0.078 at rank 25 and 95% half-widths of 0.03 to 0.06, so no cutoff by rank separates strong metrics from weak ones); and a metric listed under a signal can be the better choice for other reasons (coverage, the user's own prior). A metric's IC after removing what another metric already explains is not computed: on the same matrix it did not track the correlation (ROCE kept 59% of its IC after CurrentROIC at a correlation of 0.93), and it overstates what a second metric adds to a score that already ranks the companies almost identically.
9. Backtest Performance Metrics¶
9.1 Portfolio Simulation¶
The backtester runs a period-by-period simulation:
- On each rebalance date, compute metrics for all tickers.
- Apply filters and scoring to select a portfolio.
- Hold an equal-weight portfolio for \(\Delta\) days.
- Compute period return as the mean of individual stock returns.
- Compound the portfolio value: \(V_{t+1} = V_t \cdot (1 + \bar{r}_t)\).
- Track a SPY benchmark over the same periods.
A note on older backtest windows (survivorship bias). The database includes companies that were later delisted — bankruptcies, acquisitions, and other departures — from about 2016 onward. Earlier years mostly contain companies that survived, so backtests that reach back before 2016 look somewhat better than reality: the losers that disappeared aren't there to drag the results down. This is why the app defaults analysis windows to a 2016 start and shows a warning when you choose an earlier date. Longer windows are still useful for exploration — just treat pre-2016 results as optimistic rather than as a reliable performance estimate.
9.2 Return Metrics¶
Compound Annual Growth Rate (CAGR):
where \(T\) = number of years. A backtested CAGR above the S&P 500's long-run average (~10%) means the strategy would have beaten the market over the tested period.
Win Rate:
9.3 Risk Metrics¶
Annualized Volatility:
where \(\sigma_{\text{period}}\) is the sample standard deviation of period returns.
Maximum Drawdown:
where \(\text{Peak}_t = \max_{s \leq t} V_s\) is the running maximum portfolio value.
9.4 Risk-Adjusted Metrics¶
Sharpe Ratio:
where \(r_{\text{excess},t} = r_t - r_f / n_{\text{periods/year}}\) and \(r_f\) is the annualized risk-free rate. A Sharpe above 1.0 is generally considered good; above 2.0 is excellent. Below 0.5 means you're not being compensated well for the risk you're taking.
Sortino Ratio:
where the downside deviation uses only negative excess returns:
The Sortino ratio is preferred when the return distribution is asymmetric (positive skew), as it does not penalize upside volatility. Compare Sortino to Sharpe: if Sortino is much higher than Sharpe, your strategy has positive skew (big winners, small losers) — a desirable trait.
Calmar Ratio:
Measures return per unit of maximum drawdown — important for assessing whether the strategy's drawdown profile is tolerable. A Calmar above 1.0 means the strategy's annualized return exceeds its worst peak-to-trough drop — you're earning more than you're risking in the worst case.
9.5 CAPM Metrics¶
Beta (systematic risk):
Alpha (annualized excess return beyond CAPM):
A positive alpha indicates the strategy's returns over the tested period were not explained by market exposure alone. A positive backtested alpha means the strategy's historical returns weren't fully explained by simply riding the market up and down. A beta near 1.0 means the strategy moved roughly in line with the market; below 1.0 means less volatile, above 1.0 means more.
9.6 Other Metrics¶
Profit Factor:
A profit factor \(> 1\) means total gains exceed total losses. Values above 2.0 are strong. Think of it as "for every dollar I lose, how many dollars do I make?" A profit factor of 1.5 means $1.50 gained for every $1.00 lost.
10. Metric Computation Reference¶
10.1 Altman Z-Score¶
A composite measure of financial health:
| Symbol | Meaning |
|---|---|
| WC | Working Capital (Current Assets - Current Liabilities) |
| TA | Total Assets |
| RE | Retained Earnings |
| EBIT | Earnings Before Interest & Taxes |
| MV | Market Value of Equity |
| TL | Total Liabilities |
| Rev | Revenue |
Interpretation: \(Z > 2.99\) = safe zone, \(1.81 < Z < 2.99\) = grey zone, \(Z < 1.81\) = distress zone.
10.2 Metric Categories¶
The 150+ computed metrics fall into these categories (counts are approximate — metrics are added over time):
| Category | Examples | Count |
|---|---|---|
| Valuation | P/E, P/B, P/S, EV/EBITDA, EV/FCF, EV/Sales, FCFYield, EarningsYield — each at today / latest filing / previous filing — plus ReturnSinceFiling, GrahamNumber | 29 |
| Profitability | Gross, operating, net, EBIT, cash flow margins (current + mean), IncomeQuality | 8 |
| Returns | ROE, ROIC, ROCE, ReturnOnTangibleAssets (current, previous filing, mean), CashROA | 9 |
| Growth Rates | Revenue, EPS, FCF, OperatingIncome, margins, ROIC, Z-Score growth | 16 |
| Growth R-squared | Trend consistency (\(R^2\)) for revenue, EPS, FCF, margins, and more | 15 |
| Price Momentum | Trailing returns over 12m-1m, 6m-1m, 3m, 1m windows plus long-term reversal (60m-12m), trend consistency, distance from the 52-week high, and distance from the 200-day average (returns are size-comparable, no normalization) | 8 |
| Fundamental Acceleration | Current-vs-mean (or current-vs-previous) gaps for ROIC, ROE, operating margin, revenue, and net income | 6 |
| Volatility | Realized volatility (60-day), market beta, idiosyncratic volatility, downside deviation, largest single-day gain, and return skew | 6 |
| Earnings Quality | Accruals, Piotroski F-Score, asset growth, inventory-vs-sales and receivables-vs-revenue growth gaps, net share issuance | 6 |
| Earnings Recency | Days since the last earnings announcement and days since the last SEC filing (point-in-time, from EDGAR 8-K / report filings) | 2 |
| Cash Flow | OperatingCF, CapEx, SBC, working capital changes, buybacks, dividends | 8 |
| Yield & Shareholder | DividendYield, PayoutRatio, BuybackYield, TotalShareholderYield, NetExtFin_to_Assets | 5 |
| Safety & Leverage | Altman Z-Score, Debt/Equity, InterestCoverage, CurrentRatio, QuickRatio | 9 |
| Efficiency | DSO, DIO, CashConversionCycle, CapexToRevenue, SBCtoRevenue, R&D/Rev | 10 |
| Size & Capital | MarketCap, EnterpriseValue (each at today / latest filing / previous filing), TotalDebt, NetDebt, Equity, Intangibles | 17 |
| Per Share & Income | EPS, BookValue/Share, Revenue/Share, Revenue, NetIncome, GrossProfit | 10 |
| Technical | CurrentPrice | 1 |
A note on the Valuation category. Market cap, enterprise value, and the core multiples (P/E, P/B, P/S, EV/EBITDA, EV/FCF, EV/Sales, earnings and FCF yield) are computed in-house from trailing-twelve-month fundamentals at three prices on one consistent basis: today (the current close, the bare name), at the latest filing (the …_AtFiling suffix — priced at the close on the last 10-Q/10-K filing date), and at the previous filing (…_PreviousFiling — the prior period's TTM at the prior filing's close). The filing dates are the moment those fundamentals became public, so the snapshots carry no look-ahead; ReturnSinceFiling captures the price drift since the latest report. These replace prior vendor-sourced fields that were priced at the stale report date and used single-quarter flow (which ran roughly 4× the annualized multiple). Multiples are NaN on a non-positive denominator; yields keep their sign.
A note on the Acceleration category. Ratio-based acceleration metrics (e.g., ROIC_vs_Mean) use raw differences because ratios are already normalized across companies. Dollar-based acceleration uses two different normalizations depending on metric sign behavior: Revenue_Acceleration is a pure percentage (Current − Mean) / Mean because revenue is reliably positive; NetIncome_Surprise is scaled by market cap because net income can be negative, which makes percent-of-mean unstable.
A note on the Volatility category. These risk metrics are direction-agnostic — neither high nor low values are always preferred. Per-sector IC analysis determines the predictive sign. In Healthcare for example, low realized volatility has shown a strongly positive relationship with future returns (the classic "low-volatility anomaly"). In other sectors, higher volatility can predict higher returns when paired with reversal signals (after sharp declines, more volatile names tend to bounce harder). The group also includes market beta, idiosyncratic (market-stripped) volatility, downside-only deviation, the largest single-day gain, and return skew — the last three help flag "lottery-ticket" stocks, whose subsequent returns have been lower in the published research.
10.3 Data Quality Requirements¶
Before computing metrics for a given (ticker, date), the following minimums must be met:
- Sufficient price history available
- At least 10 income statement records
- At least 10 balance sheet records
- Positive shares outstanding
- Valid (positive) current price
Observations failing these checks are excluded from the matrix entirely.
Appendix: Why These Specific Tests?¶
The choice of statistical tools is deliberate:
| Analysis | What It Reveals | Limitation Addressed |
|---|---|---|
| IC (Spearman) | Overall rank-monotonic predictiveness, with its stability across the two halves of the window | First-pass screening of 150+ metrics |
| Quintile (held fifths) | Shape of the return-metric relationship, each fifth held as a portfolio | IC may miss non-linearities; a plain average can be carried by a few winners |
| Rule preview | What one cutoff, applied per date, would have passed and returned against the universe | A quintile is not a buy/sell rule |
| Filter funnel | What each stage of a saved strategy's filter stack removes | Single-metric tests ignore the rest of the stack |
| IC by group | Sector- and size-dependent signal strength | An overall IC can average away a group effect |
| Redundancy and independent signals | Which metrics rank the same names, and how many unrelated signals a set carries | Two correlated metrics are one bet counted twice |
By applying multiple complementary analyses, we build converging evidence about factor quality and reduce the risk of overfitting to a single statistical artifact.
Glossary¶
| Term | Definition |
|---|---|
| IC (Information Coefficient) | A measure of how well a metric's rankings predict future stock returns. Higher absolute values mean stronger predictive power. |
| Spearman correlation | A rank-based correlation that measures whether two variables move together in order (not necessarily in a straight line). More robust than standard (Pearson) correlation for financial data. |
| Quintile | One of five equal-sized groups when all stocks are sorted by a metric. Q1 is the bottom 20%, Q5 is the top 20%. |
| CAGR (Compound Annual Growth Rate) | The smoothed annual return that gets you from the starting value to the ending value. Accounts for compounding. |
| Sharpe Ratio | Return per unit of total risk (volatility). Higher is better. Above 1.0 is good; above 2.0 is excellent. |
| Sortino Ratio | Like Sharpe, but only penalizes downside volatility. Rewards strategies with big gains and small losses. |
| Alpha | The portion of a strategy's return that can't be explained by market movements. Positive alpha = the strategy adds value beyond just tracking the market. |
| Beta | How much a strategy moves relative to the market. Beta of 1.0 = moves with the market. Below 1.0 = less volatile. Above 1.0 = more volatile. |
| Max Drawdown | The largest peak-to-trough decline in portfolio value. Measures worst-case pain. |
| p-value | The probability that a result as extreme as the observed one could occur by chance. Below 0.05 is conventionally considered statistically significant. |
| Forward return (FutureProfit) | The actual stock return over the next rebalance period (default 90 days). This is what every analysis is trying to predict. |
Further Reading¶
- Product Overview — Feature summary and getting started
- Screen Guide — Detailed walkthrough of every screen in the app
- Setup Guide — Installation, IBKR configuration, and first-launch walkthrough