Quantitative finance glossary

One half of the sum of absolute portfolio-weight differences from a benchmark.

Dimensionless fraction or percentage of NAV.

From 0 to 1 for standard fully invested long-only portfolios; it can exceed 1 with leverage or short positions.

Zero means identical weights; larger values mean a greater holdings-level departure from the benchmark.

It measures portfolio difference directly and complements return-based tracking error.

A measurable security characteristic or signal used to forecast relative future returns.

Input-specific before normalization; ranks, percentiles, and z-scores are dimensionless.

Depends on construction. A z-score is unbounded; a percentile is usually in [0,1] or [0,100].

The direction convention must state whether larger values predict larger or smaller subsequent returns.

Factor definition, timestamp, normalization, and universe determine the evidence evaluated by a backtest.

Uncertainty in an expected-return estimate, represented by completed forecast errors, an uncertainty set, or a probability model.

Expected return over the forecast horizon. Forecast, error covariance, and portfolio weights must use compatible units.

Error variance is non-negative; correlated errors can increase or reduce portfolio-level uncertainty depending on the positions.

Greater uncertainty lowers the robust value assigned to an otherwise identical expected-return forecast.

It prevents an optimizer from treating small, imprecisely estimated differences in expected return as exact inputs.

A portfolio construction target in which estimated market beta is approximately zero.

Dimensionless, provided asset and market returns use the same currency and return convention.

Target is zero within a stated tolerance. Realized beta can move as holdings and covariances change.

The portfolio is designed to have little first-order sensitivity to the specified market factor.

It attributes returns after controlling for broad market direction. Sector, currency, and nonlinear risks remain in the portfolio.

The amount of capital a strategy can deploy before market impact, liquidity, or crowding materially reduces its expected performance.

Capital in a declared base currency, such as USD, EUR, GBP, or JPY; model uncertainty may warrant reporting a range.

Non-negative and model-dependent. Upper bounds vary with liquidity, impact assumptions, and the required net return.

It is the deployable scale consistent with a stated net-performance or market-impact threshold.

Capacity determines how much capital can earn the modeled net return.

The combination of a responsive covariance estimate with a more stable target to reduce estimation error and improve numerical conditioning.

The estimate and target must use the same assets, return horizon, currency, annualization, and covariance units.

A weight of zero retains the responsive estimate; a weight of one uses the target. Intermediate values trade responsiveness for stability.

More shrinkage is useful when the responsive estimate is noisy, provided the target contains relevant structure.

Portfolio optimization can amplify small covariance errors, particularly along low-variance eigenvectors.

Ordering securities against one another at the same observation time using a signal or characteristic.

Rank is an integer; the displayed percentile rank q is dimensionless.

Ranks run from 1 to N_t and the displayed convention maps them to [0,1]. Other percentile conventions and ties must be stated.

It describes a signal’s relative position within the eligible universe. Economic magnitude requires a separate measurement.

Ranking reduces sensitivity to outliers. Results depend on universe composition, tie handling, and direction.

A factor held through many correlated positions among market participants, creating exposure to fast, correlated unwinds alongside its average historical return.

Days to cover is expressed in trading days; comomentum and ownership-concentration proxies are dimensionless.

Higher short interest, days to cover, or comomentum indicate more crowding; there is no universal threshold.

Crowding describes positioning and correlated-unwind risk independently of expected-return direction or magnitude.

Greater crowding increases exposure to sudden, correlated drawdowns for a given level of expected return.

A portfolio whose long and short base-currency notionals offset, producing approximately zero net notional exposure.

Notional in the portfolio base currency b; X^{b/l} is units of base currency per unit of local currency.

Net notional targets zero within tolerance; gross notional remains positive and must be reported separately.

Long and short market values offset after FX translation. The conventional name applies even when b is EUR, GBP, JPY, or another currency.

It controls net capital exposure; beta, factor, and FX exposures require separate constraints.

A portfolio whose weights are derived from an eigenvector of an asset return covariance or correlation matrix.

Eigenvalues have return-variance units for covariance PCA and are dimensionless for correlation PCA; weights require a declared normalization.

Eigenvalues of a valid covariance matrix are non-negative. Eigenvector elements can have either sign, and the vector sign is arbitrary.

Each eigenportfolio represents an orthogonal sample direction of return variation, ordered by its associated eigenvalue.

Eigenportfolios expose dominant covariance directions, support statistical factor models, and help diagnose hidden portfolio concentration.

The conditional return estimate supplied to an allocation model. Realized returns can differ from the estimate.

Return over horizon h or annualized return, in a declared base currency b. The calculation must state whether it uses arithmetic or log returns.

A simple-return forecast is bounded below by −100% and unbounded above; model estimates may require clipping or shrinkage.

The estimate conditions on information available at t and applies to the stated forecast horizon.

Optimized weights can be highly sensitive to small differences in expected-return estimates.

The number of independent investment decisions made by a forecast process over a stated measurement period.

Independent decision count per stated period, such as per rebalance or per year.

Non-negative. A value near one behaves like one independent decision; larger values indicate more distinct opportunities.

Correlation across securities, signals, holdings, and overlapping decision times reduces effective breadth below the raw opportunity count.

Breadth determines how repeated modest forecast skill can accumulate and prevents universe size from being mistaken for independent information.

The sensitivity of a security or portfolio to a specified factor, estimated from holdings, characteristics, or a return model.

Regression beta is dimensionless when portfolio and factor returns use the same units; characteristic exposures retain their stated scale.

Generally unbounded. Zero means neutral to the specified factor model; residual, idiosyncratic, and omitted-factor risks remain.

Positive and negative values describe the direction and magnitude of sensitivity under the chosen model.

Unintended exposures can make apparently different strategies respond to the same underlying risk.

The removal of stated common exposures from a security score or portfolio through residualization or explicit holding constraints.

Residual scores retain the scale implied by the input and regression; portfolio exposure constraints use the units of the exposure matrix.

A zero residual equals the value predicted by the included exposures. Portfolio neutrality targets zero within a stated tolerance.

Positive and negative residuals lie above and below the component explained by the specified exposures.

Neutralization reveals whether signal evidence remains after sector, size, market, or other common relationships are controlled.

The deliberate variation of factor exposure through time using forecasts of factor performance or risk.

Factor weights are fractions of capital or risk; predictors retain their declared units until normalized.

Defined by the timing rule and exposure bounds; zero means the factor is omitted at that date.

Exposure changes because the expected opportunity or risk is believed to vary through time.

Timing adapts exposure to regimes and adds estimation error, turnover, and another layer of model-selection risk.

A portfolio of security weights constructed to represent a chosen factor while controlling declared common exposures and implementation limits.

Weights are fractions of capital or gross notional; factor exposure uses the scale defined by the signal and construction.

Long and short weights can be positive or negative. Net, gross, target exposure, and risk normalization must be reported separately.

The resulting return series records how the factor behaved after security covariance, neutralization, constraints, and implementation rules.

It turns a security-level score into a controlled portfolio whose return, risk, turnover, drawdown, and covariance can be evaluated.

A strategic approximation relating potential active Information Ratio to forecast skill, independent breadth, and forecast transfer into positions.

Dimensionless. IC, transfer, and breadth must describe compatible forecast horizons and measurement periods.

The sign follows IC times transfer; breadth is non-negative and enters through its square root.

Potential risk-adjusted active performance rises with repeatable forecast skill, independent opportunities, and stronger portfolio transmission.

The decomposition helps locate whether a process is limited by its signal, duplicated decisions, or restrictive implementation.

The share of completed evaluation dates on which a factor’s date-level information coefficient has the registered positive sign.

Dimensionless proportion or percentage. The report must state its event, observation frequency, zero rule, and valid date count.

From 0% to 100%. A higher value means positive IC occurred more often; it contains no information about the size of hits or misses.

It complements mean IC by separating the frequency of the intended sign from the magnitude of each date-level correlation.

Two factors can have opposite rankings by hit rate and mean IC, so both the count and full IC distribution belong in evaluation.

The performance difference between a paper portfolio at the decision price and the realized executed portfolio.

Return, percentage, basis points, or base-currency amount. All legs and fees must be translated to the same base currency b.

Signed; positive commonly denotes a cost. It can exceed the quoted spread because it includes delay, impact, fees, and missed trades.

It measures the total economic gap between an investment decision and its implementation.

It combines slippage, delay, impact, fees, and missed trades to connect paper performance with realized performance.

The cross-sectional correlation between a signal score and a subsequent return. Rank IC uses ranked values and measures whether the signal orders securities correctly.

Dimensionless. Returns must use one horizon and one base currency b across the cross-section.

From −1 to +1. Zero indicates no monotonic cross-sectional association.

The sign gives ranking direction; magnitude gives the strength of association on one observation date.

It separates signal-ranking evidence from the realized performance of a particular portfolio construction.

The change in a signal’s information coefficient as its forward-return endpoint moves farther from the signal observation time.

IC is dimensionless; the horizon must state trading periods, calendar time, or another explicit clock.

The curve may rise, peak, decay toward zero, or reverse sign. Every point requires the same signal, universe, return, currency, and IC conventions.

It identifies when a fixed signal snapshot has its strongest measured association with cumulative or incremental future returns.

Forecast timing informs later choices about signal refresh, holding periods, turnover, and cost-aware portfolio tests.

The absolute size of a date-level or mean information coefficient, interpreted with its convention, horizon, universe, sampling history, and implementation setting.

Dimensionless. The report must distinguish a single-date IC from a mean estimated across dates.

There is no universal cutoff. Larger absolute values indicate stronger measured association under the stated specification, while persistence and uncertainty require the full IC history.

Magnitude becomes practically useful when the relationship repeats out of sample and survives breadth, portfolio constraints, turnover, and costs.

A fixed numerical threshold can misstate evidence when the universe, horizon, dependence, or implementation changes.

The degree to which an estimated information coefficient is supported by a fixed specification, a repeatable history, and uncertainty calculations that reflect cross-sectional and time-series dependence.

IC and its interval are dimensionless. The report should state the number of dates, valid security counts, confidence level, and dependence adjustment.

A narrower interval indicates greater precision under the stated sampling model. An interval spanning zero does not establish the registered positive or negative direction at that confidence level.

Reliability depends on the research design, serial and cross-sectional dependence, sample length, and any selection across alternative specifications.

It prevents a point estimate from carrying more evidential weight than its observation history and sampling design support.

The mean return of a portfolio relative to a declared benchmark divided by the standard deviation of that active return, usually annualized.

Dimensionless. Portfolio and benchmark returns must use the same period and base currency; the report must state the benchmark, annualization factor, and gross or net treatment.

Unbounded in principle. Positive values record positive mean active return; a zero tracking-error denominator makes the ratio undefined.

It measures benchmark-relative portfolio return per unit of realized active-return variability over the stated sample.

It evaluates the portfolio result after forecasts have passed through position sizing, constraints, trading, and implementation costs.

The largest peak-to-trough decline in a cumulative return or portfolio-value series.

Dimensionless fraction or percentage of portfolio value in the reporting base currency.

From 0 to 100% for a non-negative unlevered wealth series; leveraged strategies can lose more than initial capital.

It reports the worst historical loss from a prior high before recovery, if any.

Drawdown captures path-dependent capital loss that volatility and Sharpe ratio can obscure.

Factor inputs or scores that are unavailable or invalid under the point-in-time timing, eligibility, and data-quality rules.

Availability is a binary indicator; coverage is a dimensionless fraction or percentage of the eligible universe.

Coverage ranges from 0% to 100%. A missing value remains distinct from a measured zero even when neutral calculation treatment is used.

Missingness can reflect reporting schedules, invalid denominators, inapplicable fields, or operational failures and should retain its cause.

Exclusion, imputation, neutral treatment, and partial composites change the evaluated sample and the effective factor definition.

The use of two or more covariance estimates to evaluate the same portfolio through a blend, separate penalties, separate limits, or a worst-model rule.

Return volatility over a stated horizon, commonly annualized. Every model must use compatible assets, currency, frequency, and scaling.

Each predicted risk is non-negative. Model disagreement is the difference between estimates for the identical position.

Separate limits control every included estimate, while covariance blending controls their weighted average.

Competing models can reveal concentrations hidden by a single estimated covariance matrix.

The selection of portfolio weights to maximize or minimize a stated objective subject to constraints.

Weights are fractions of capital or gross notional. Every objective term must be placed on a consistent return and time scale.

The feasible values are defined by the budget, bounds, leverage, liquidity, currency, and exposure constraints in the set W.

The specified inputs, objective, constraints, and numerical method determine the result.

It converts forecasts and risk estimates into a capital allocation under explicit assumptions.

The part of an expected-return forecast that remains after the component associated with specified common exposures has been removed.

Expected return over a declared horizon and base currency, or a dimensionless score after normalization.

Positive and negative values lie above and below the forecast associated with the included exposures; zero equals the fitted common component.

Residual alpha ranks securities relative to what their market, sector, style, or other included exposures would predict.

It separates a security-selection forecast from common relationships that can otherwise dominate raw expected return.

A limit or allocation assigned to a source of portfolio risk, such as an asset, strategy, or factor.

Risk contribution has the same volatility units as the portfolio, commonly annualized percentage points; normalized contributions are percentages.

Normalized contributions commonly sum to 100%; individual contributions can be negative when a position hedges other risk.

It expresses each position’s or factor’s marginal contribution to portfolio volatility.

Differences in volatilities and correlations often give equal capital weights unequal risk contributions.

The mean information coefficient divided by its standard deviation over a trailing window, usually annualized. It measures the persistence of ranking skill across observations in the window.

Dimensionless. A is observations per year, commonly 252 for daily non-overlapping observations.

Unbounded in principle. Positive is favorable for the stated signal direction; instability increases when the denominator is small.

A high value requires positive average rank association that is reasonably consistent through the trailing window.

It is useful for monitoring signal decay and for distinguishing persistent evidence from a few unusually strong dates.

The cross-sectional correlation between signal and forward-return ranks after a declared sector or industry component has been removed from both rank fields.

Dimensionless. The report must state the point-in-time classification, group level, weighting, fields adjusted, operation order, and minimum group size.

From −1 to +1. Its sign and magnitude describe residual or within-group ordering under the stated neutralization method.

Comparison with raw IC shows how much measured association came from within-sector company ordering and how much came from differences between sectors.

It aligns factor evaluation with a within-sector research claim or a portfolio mandate that tightly controls sector exposure.

Mean excess return divided by return volatility, expressed on a consistent time scale.

Dimensionless. The return frequency, annualization factor A, base currency b, and risk-free convention must be stated.

Unbounded in principle. Positive means positive average excess return; comparisons require consistent sampling and costs.

It measures average excess return per unit of total realized volatility.

Report drawdown, tail risk, and estimation uncertainty alongside this risk-adjusted summary.

The signed difference between a reference price and the achieved execution price.

Dimensionless return, usually basis points. Prices and FX conversion must use a consistent quote convention and timestamp.

Signed and unbounded in theory; positive commonly denotes a cost when side is +1 for buys and −1 for sells.

It isolates price deterioration or improvement relative to the selected reference price.

Small per-trade differences can consume a large share of gross alpha in high-turnover strategies.

Mean return above a minimum acceptable return divided by downside deviation.

Dimensionless. Returns, minimum acceptable return, time scale, and base currency must be consistent.

Unbounded in principle and unstable when there are very few downside observations.

Its denominator uses returns below the chosen threshold and excludes upside variability.

It focuses the risk adjustment on harmful downside variation.

The standard deviation of active return relative to a benchmark.

Return per year, normally reported in annualized percentage points. Portfolio and benchmark must share base currency b.

Non-negative. Zero means identical measured returns; portfolios with different holdings can produce that outcome.

Higher tracking error means realized performance deviates more widely from the benchmark.

It connects active portfolio decisions to the variability of benchmark-relative outcomes.

A dated, reproducible rule that converts information available at a decision time into a forecast or action-relevant score.

Input-specific before normalization; the output scale and observation cutoff belong to the versioned definition.

The domain depends on the rule. Positive, zero, and negative values require a recorded orientation and forecast target.

One value is a signal observation; applying the same rule across a universe and through time produces the factor record used in evaluation.

Separating the observation, factor definition, and portfolio decision prevents a current score from being mistaken for a trade or a performance result.

The correlation between risk-adjusted return forecasts and the risk-adjusted active positions produced after portfolio constraints.

Dimensionless correlation under a declared benchmark and risk-model convention.

From −1 to +1. One is perfect positive alignment, zero is no linear transfer, and negative values oppose the forecasts.

A lower value records how sector, factor, long-only, turnover, position, or other constraints change forecast-aligned positions.

It separates forecast skill from the efficiency with which portfolio construction carries the forecasts into holdings.

The amount of portfolio trading required between two sets of weights, under a stated convention.

Fraction or percentage of NAV/gross notional per rebalance; it may also be annualized. The one-way or two-way convention must be stated.

Non-negative. It can exceed 100% over a period or at one rebalance for leveraged portfolios.

One-way turnover of 0.25 means trading notionals equal to roughly 25% of the chosen capital base.

Turnover connects weight instability to commissions, spreads, market impact, taxes, and operational load.

The point-in-time rules determining which securities are eligible for research or portfolio construction.

Binary indicator for each security and observation date; aggregate universe size is a security count.

0 for ineligible and 1 for eligible. Membership must be reconstructed point in time.

The universe defines the cross-section against which signals are ranked and portfolios are formed.

Survivorship, liquidity, listing, country, and data-availability rules can materially change measured factor performance.

A transformation that retains tail observations while replacing values below and above declared thresholds with those thresholds.

The capped value keeps the raw field units; the threshold rule and calculation universe must be recorded.

All outputs lie between the lower cap L and upper cap H, inclusive.

Tail observations keep their cross-sectional membership but no longer retain their full raw distance from the rest of the universe.

It can limit the influence of valid extremes on normalization and holdings, but it should not be used to conceal invalid data.

A standardized value expressed as its distance from the cross-sectional mean in estimated standard-deviation units.

Dimensionless. The universe, weighting, dispersion estimator, and treatment of zero dispersion must be stated.

Unbounded in principle. Zero equals the estimated mean; positive and negative values lie above and below it.

Magnitude retains estimated raw distance from the mean and can therefore remain sensitive to skewness and extreme observations.

Z-scoring makes unlike input units comparable while preserving more magnitude information than cross-sectional ranking.