The LRT Alpha Framework connects several distinct analytical disciplines. Defined factor series describe common return patterns. A joint model estimates how individual assets participate in those patterns. Measurements of exposures and residual behavior become inputs to LRT Alpha Score, LRT Capital's proprietary composite measure of relative asset attractiveness. Portfolio construction then considers Alpha Scores alongside risk estimates and constraints, while a separate research process evaluates the relationship with subsequent returns.
LRT Alpha Score combines proprietary inputs into a relative asset ranking. The broader explanatory model and the inputs selected for the score serve different purposes: the model describes return behavior; Alpha Score expresses the resulting assessment.
The LRT investment process
LRT's investment process connects three decisions: which businesses qualify for investment, which eligible stocks offer stronger expected returns, and how those opportunities should be combined into a portfolio.
The first decision establishes the foundation. We seek businesses with durable competitive advantages, attractive growth opportunities and management teams that allocate capital intelligently. Passing this assessment earns a company a place in the opportunity set. The next decisions concern its expected return and its role in the portfolio.
Our process begins with business quality, uses LRT Alpha Score to assess relative return opportunity, and applies portfolio construction to manage the exposures taken in pursuing those returns.
| Step | The question | Focus | Result |
|---|---|---|---|
| 1. Define the universe | Which businesses are we willing to own? | Competitive advantage, growth opportunities and management capital allocation | An eligible opportunity set |
| 2. Forecast expected returns | Which eligible stocks offer stronger return opportunities? | LRT Alpha Score and a consistent comparison within the universe | Relative expected-return assessments |
| 3. Construct the portfolio | How should we allocate capital and manage the resulting exposures? | Position sizing, factor balance, concentration controls and index hedges | A portfolio tilted toward stronger forecasts with deliberate risk exposures |
What earns a business a place in the universe?
The opportunity set is a deliberate investment choice made before expected-return forecasts and position sizes. We examine four connected questions.
What protects the business from competition? Customer loyalty, switching costs, scale, networks, intellectual property and distribution can help protect a company's economics. We look for the mechanism behind persistent pricing power, attractive returns on invested capital and resilient customer relationships. A talented executive or a successful product alone does not establish a moat. We also ask how technology, competition or changing customer behavior could weaken the protection, and which customer needs are likely to endure.
Where can the business reinvest at attractive returns? Growth creates value when the economics of expansion justify the capital required. We examine the runway for reinvestment, whether the competitive advantage extends to new customers or adjacent activities, and what shareholders receive as the business grows. Revenue growth by itself does not establish value creation; the return on additional capital matters.
How does management allocate capital? Reinvestment, acquisitions, financing, dividends, repurchases and share issuance all affect value per share. We examine acquisition discipline, incentives, governance and the cash consequences of past decisions. Management can strengthen a moat by investing within the company's advantage or weaken the economics through unrelated expansion and poorly priced acquisitions. A strong operating business still needs intelligent capital allocation to translate its cash generation into shareholder value.
Can we understand the cash economics and financing? The qualitative review considers cash generation, financial transparency, governance and debt relative to the stability of the business. An appealing growth story is less useful when the cash economics cannot be followed or financing makes the company fragile. These are investment judgments, rather than a mechanical screen using one accounting ratio.
Eligibility establishes which businesses we are willing to consider. It does not determine which eligible stock offers the strongest expected return or deserves the largest position today. The factor-modeling stages below support those subsequent return and exposure decisions.
Can you predict stock returns?
A useful return forecast makes a statement about expected opportunity before the outcome is known. An expected return describes a central expectation across possible outcomes over a stated horizon; the actual return can land above or below it. It can help us compare assets, estimate the size of a return difference, and decide how to express those views in a portfolio. Individual realized returns still contain substantial uncertainty.
The LRT Alpha Framework follows a logical sequence. Model the shared influences on returns. Measure how each asset participates in those influences and how its remaining behavior differs. Combine selected measurements into Alpha Score. Study whether stronger scores are associated with better subsequent outcomes. Then bring the return assessment together with factor exposures, co-movement and portfolio constraints.
Forecast quality must be evaluated against a stated target. Ranking assets, predicting the return difference between diversified groups and predicting an individual stock's exact return are different tasks. The research below measures the score/return relationship. A separate chronological basket exercise, explained in the investor guide, studies forecasts of group differences. Neither statistic is a live managed-portfolio track record.
One connected process, with distinct decisions
The analytical depth comes from how the stages work together. Factor construction determines what is measured. Joint estimation determines how the measurements relate to each asset's returns. Standardization makes the selected inputs comparable. Signed scoring weights determine which combinations are favored. Portfolio rules determine how Alpha Scores can be expressed within the chosen risk framework.
Research provides another view of Alpha Score: compare it with returns that occur afterward, vary the statistical assumptions and inspect whether the relationship depends on a narrow subset of observations. The Alpha Score study follows this research branch. It evaluates the score before attributing an outcome to a particular portfolio implementation.
This separation makes the process easier to examine. A change in a factor's definition is different from a change in an asset's fitted exposure. A change in an Alpha Score weight is different from a change in a portfolio limit. Each has a defined place in the architecture and a different interpretation.
Where the analytical depth comes from
Factor engineering creates a defined measurement
A factor series is produced by a specified construction. Depending on the input, that can involve a reference return series, ranked security baskets, adjustments for overlapping influences and a chosen scale. The definition, weighting convention and update rule establish what the resulting series represents.
This stage turns a broad investment idea into observations that can be used consistently across assets and dates. The detailed construction belongs upstream of the final Alpha Score; it is not inferred from a favorable ranking chart.
Joint estimation examines the influences together
The return model fits multiple explanatory series together. This matters because an asset can participate in several overlapping influences at once. Regularization controls the size of estimated coefficients and can leave some coefficients at zero, rather than assigning a separate large contribution to every available input.
The result is a description of an asset's sensitivities conditional on the fitted model. It provides a richer measurement framework than treating every factor relationship as an independent comparison.
The model produces two useful sets of measurements
One set describes the asset's fitted exposures. Another describes the residual behavior remaining after the model's fitted return is removed. These are distinct inputs to analysis: common influences explain part of the observed behavior, while the remaining behavior can be summarized separately.
Selected measurements from these channels enter the Alpha Score calculation. A scoring input can summarize an exposure or an aspect of residual behavior; it does not have to be a traded factor basket. The model's explanatory series and the score's selected inputs should not be treated as interchangeable.
A common scale connects measurement to Alpha Score
The selected inputs are standardized across the comparison universe and combined using signed weights. This makes each input's direction within Alpha Score explicit and prevents differences in raw measurement units from determining the score by accident.
Recent fitted measurements are smoothed before scoring. Sparse zero coefficients remain part of the average over fitted sessions; unavailable fits are handled separately. Those details determine whether the implemented score represents the intended calculation consistently.
Portfolio construction adds a separate decision layer
A high Alpha Score alone does not determine a position size. Portfolio construction brings in estimated co-movement between assets, exposure requirements and position limits. Alpha Score expresses the relative asset ranking; the construction stage considers how those rankings fit together under the chosen risk and constraint settings.
This distinction allows a single analytical framework to connect asset-level measurements with portfolio-level decisions without treating the two as interchangeable outputs.
Research checks the relationship and the calculation
The study evaluated repeated monthly observations, several forward horizons, ranked groups and alternative specifications. It also reconstructed scores independently and compared them with the source calculation: the checked asset-score comparisons matched to floating-point precision.
The return comparisons and the calculation checks answer different questions. The former examine the historical score/return relationship. The latter check that the research score corresponds to the production formula. Together they make both the empirical result and the underlying calculation more open to examination.
What a factor measures
A factor is a defined source of common variation. It can be represented by a reference return series or by the difference between two groups of securities selected according to a characteristic. A factor gives researchers something measurable to compare with an asset's returns over time.
Three ideas are especially useful. A descriptor measures a characteristic at a particular date. A factor return measures how a reference series or a constructed basket behaves during a period. A factor exposure, often called a loading, measures the historical sensitivity of an asset to that return series.
These numbers answer different questions. A descriptor asks, “What does this asset look like?” A factor return asks, “How did this common influence behave?” An exposure asks, “How much has this asset tended to move with that influence?” Keeping them separate makes the whole process easier to understand.
A positive exposure indicates that, within the fitted model, an asset tends to respond in the same direction as the factor. A negative exposure indicates the opposite direction. The magnitude depends on the units of the factor and the model specification. An exposure is a historical sensitivity, not an automatic forecast of the factor's next return.
How factors are created
1. Define an economic or statistical idea
Factor construction starts with a precise question. Does a particular company characteristic distinguish groups of assets? Does an observable reference series help describe common movement? Can a recurring pattern be expressed as a transparent calculation?
The definition needs to specify the eligible securities, the measurement, the direction of the comparison and the update schedule. A vague investment story becomes testable only after those decisions are translated into rules.
The examples here explain general construction techniques. Our interactive factor explorer publishes definitions, construction notes and downloadable return data for the factors shown there. Alpha Score's input selection, scoring weights and implementation settings remain proprietary.
2. Prepare the observations
The next step is to align the underlying data. Price observations need a consistent adjustment convention. Company information needs a usable observation date. Securities need comparable units and identifiers. The intended timing matters: an item describing an earlier reporting period is not necessarily usable before its publication date.
Researchers also decide how to handle extreme measurements, missing information and unusual securities. Depending on the design, a measurement can be ranked, expressed as a percentile, scaled relative to peers or compared with an asset's own historical range. These are alternative construction choices rather than a claim that every transformation is used for every factor.
3. Form a return series
One approach uses an observable reference series directly. Another sorts eligible assets by a characteristic and creates two baskets: one representing one end of the characteristic and another representing the other end. A simple factor return is the first basket's return minus the second basket's return.
For example, if a fictional high-characteristic basket gains 2.0% and a fictional low-characteristic basket gains 0.5%, their return difference is 1.5 percentage points. If both baskets fall, the difference can still be positive when the first falls less than the second.
Basket membership and weighting rules define what this spread measures. Equal weighting and capitalization weighting answer different questions. Rebalancing refreshes the baskets as the underlying characteristics change. A calculated spread is a research series; translating it into a traded position introduces financing, execution and other implementation considerations.
4. Separate overlapping influences
A basket built around one characteristic can also move with broader influences. Researchers can estimate that overlap and remove a fitted component, producing a more targeted factor series. This process is often called neutralization or orthogonalization.
The aim is to describe the chosen influence more distinctly. It does not make a factor economically isolated or permanently uncorrelated with everything else. The adjustment depends on the comparison factors, the estimation window and, in a sequential construction, the order of the calculations.
5. Put the series on a consistent scale
Different factor series can have very different natural amplitudes. Scaling can make them easier to compare within a model. A volatility-scaled series, for example, changes the size of its daily movements to a chosen reference scale using an estimate of historical variability.
Scaling changes the units of both factor returns and the exposures fitted against them. It should therefore be part of a documented definition. A large numerical loading on one scale is not directly comparable with a loading on a differently scaled factor.
6. Maintain the definition through time
A useful factor is more than a one-time sort. It needs a repeatable process for refreshing data, updating baskets, calculating returns and preserving observations for later analysis. That history allows researchers to ask how assets behaved relative to the factor at different dates.
The practical discipline is consistency: apply the specified calculation, retain its timing convention and distinguish a change in the factor's construction from a change in market behavior.
A factor in action
The interactive example below shows the simplest basket-spread calculation. Change the returns of two fictional groups to see how their difference changes. The values are invented for explanation and have no connection to the study's assets, proprietary factors or results.
The subtraction matters more than either basket in isolation. If both rise by the same amount, the factor spread is zero. If the first group outperforms the second, the spread is positive. If it underperforms, the spread is negative. This is one way to turn a relative characteristic into a measurable return series.
From factor returns to an asset's exposure
Once factor return histories exist, the model compares them with an individual asset's return history. A simplified relationship is:
Asset return ≈ fitted baseline + factor contributions + residual
Factor contribution ≈ fitted exposure × factor return
The model estimates several exposures together. An asset can participate positively in one influence, negatively in another and only weakly in a third. This collection of sensitivities describes more than a single broad measure of market movement.
For a fictional one-period example, suppose a reference factor moves by 1% and an asset's fitted exposure is 0.8. The corresponding fitted contribution is 0.8 percentage points. An exposure of -0.8 would produce a contribution in the other direction. Actual return can differ because the other factors, fitted baseline and residual also contribute.
The residual is the difference between the observed return and the model's fitted return. It can contain company-specific events, omitted influences and estimation error. It is a model-dependent remainder, not automatically a measure of skill or a forecast. Its behavior over time can nevertheless be summarized and studied as an additional source of information. Some scoring inputs describe factor exposures; others describe characteristics of that remaining behavior.
Why estimate the factors together?
Several factor series can overlap. Looking at each separately can attribute the same part of an asset's movement to multiple influences. Joint estimation examines the relationship while accounting for the other series included in the model.
The research framework uses regularized joint estimation. Regularization discourages unnecessarily large or unstable coefficients and can reduce some coefficients to zero. This is useful when explanatory series overlap and ordinary regression would be sensitive to small changes in the observations.
A zero coefficient means the fitted specification assigned no separate contribution to that input in the estimation window. It does not establish that the underlying economic influence never matters. Similarly, regularization is an estimation tool; its use alone does not prove that a model predicts future returns.
In the stored fits used for the study, asset-return observations precede the relevant fit date. Exposures are then smoothed over recent fitted sessions before they enter the Alpha Score calculation. This produces a sequence of measurements that can be compared at successive observation dates.
Turning measurements into Alpha Score
The next step is to put comparable measurements on a common scale. A raw exposure with a large numerical value should not dominate simply because its underlying factor uses different units.
The scoring process standardizes each selected input across the comparison universe. A standardized value describes where an asset sits relative to other assets at the same observation date. The selected inputs are then combined using signed Alpha Score weights:
Alpha Score = sum of the selected inputs' weighted contributions
The sign of a scoring weight determines which direction of an input is favored. A positive exposure is not inherently a favorable reading: it can increase or decrease Alpha Score depending on the weight assigned to it. Several contributions can reinforce or offset one another.
Alpha Score is relative to the comparison universe. A higher value expresses greater relative asset attractiveness within that universe; it is not a percentage-return target or a price forecast. When the universe or the saved scoring weights change, the resulting comparisons can change too.
For this study, the composite is standardized once more among the assets with usable scores in a month. That makes a one-standard-deviation score difference an interpretable unit for comparing the regression estimates across observation dates.
Why refresh the measurements?
Assets do not have permanently fixed exposures. Their businesses, trading patterns and relationships with common influences can change. Refreshing the fitted measurements allows the description of an asset to evolve with the data used by the model.
Smoothing recent usable fits reduces the impact of isolated daily changes. It averages fitted observations rather than pretending an unavailable fit is another valid day. A coefficient that the estimator sets to zero still belongs in that fitted-session average.
The study samples Alpha Score at monthly anchors. Monthly sampling is a research convention used to compare a sequence of observations with subsequent outcomes; it is not itself a statement that every input is constructed monthly or that a live portfolio must trade at that frequency.
A score is one part of an investment process
Alpha Score helps organize the relative attractiveness of assets. Position sizing must also consider how holdings interact, the amount of risk to take, portfolio constraints, concentration, liquidity and the costs of implementing a trade.
Those decisions are related to Alpha Score but are not contained in the score alone. Two portfolios using the same Alpha Score ranking can produce different outcomes because they size positions differently or operate under different constraints. The research presented here evaluates Alpha Score at monthly observation dates against later asset returns, before translating it into a particular managed portfolio.
From return forecasts to portfolio weights
All else equal, a stronger expected return supports a greater portfolio weight. In practice, that return view is evaluated alongside forecast uncertainty, estimated covariance, factor exposures, position limits, liquidity and trading costs. The objective is to express the favorable return assessments in a combination of positions whose shared risks are understood.
Portfolio factor exposure is the weighted combination of its positions' exposures. Adding companies does not necessarily diversify that exposure: apparently different stocks can depend on the same influence. We therefore examine both individual positions and the return patterns they contribute to the portfolio.
A lower-forecast stock can still play a useful role if it reduces an important concentration or provides a sufficiently different return pattern. Its contribution must be assessed in the context of the whole portfolio. The trade accepts less standalone expected return from that position in exchange for a potential reduction in shared risk. The attractiveness of that exchange depends on the estimates and constraints; diversification alone is not a reason to buy every lower-forecast asset.
This explains how a portfolio can favor stronger forecasts while retaining some lower-ranked positions. The portfolio's return tilt and its diversification requirements are considered together. The resulting weights need not reproduce the Alpha Score ranking in strict order, and factor diversification does not require neutralizing every exposure.
A long history across changing conditions
The research uses a long historical period with complete observations across different market settings: the recovery following the global financial crisis, near-zero interest rates and the COVID disruption, the inflation surge and rapid monetary tightening, and a U.S. sovereign credit-rating downgrade.
These events describe the backdrop of the available sample. The reported estimates pool eligible observations rather than establishing separate forecasting results for each event. The completed score history covers the financial-crisis recovery, not the earlier crisis selloff. Monthly anchors with unresolved historical fits were excluded in full; missing fits were not replaced by invented scores.
Absolute returns, abnormal returns and surprises
An asset can earn a positive return because the broad market rose, because its systematic exposures were rewarded, because of asset-specific behavior, or through some combination of these influences. The return definition determines which part of that outcome the research measures.
Absolute return is the asset's exit price divided by its entry price, minus one. The prices used here already incorporate the database's corporate-action adjustments. Those price adjustments are separate from risk adjustment: an adjusted-close return can still be an absolute return with all of the asset's market exposure included.
Benchmark-relative return subtracts a common market benchmark's return over the same dates. It measures performance relative to that reference but does not account for different assets having different exposures. Within a monthly cross section, subtracting the same benchmark from every asset changes the intercept, but not the Alpha Score slope or the highest-minus-lowest group spread.
Exposure-adjusted abnormal return subtracts a model-based normal holding-period return. The primary models use a short-term Treasury-fund cash proxy plus contributions from exposures estimated strictly before the score date. One model controls for broad market exposure; another uses a broader set of return factors. Their coefficients are estimated separately from the proprietary score calculation.
The normal daily returns are compounded across the holding period before comparison with the asset's actual return. The calculation uses the factor returns realized during that holding period for attribution. It does not use those future factor returns to construct Alpha Score or estimate its prior exposures.
Normal daily return = cash reference + fitted factor contributions
Normal holding-period return = product(1 + normal daily return) − 1
Abnormal return = absolute return − normal holding-period return
The primary abnormal-return reference excludes the asset's fitted historical intercept. A separate historical-model surprise test uses a fuller fitted normal return, including an asset-specific baseline called an intercept. This changes the question from “what remains after allowing for factor exposures?” to “what remains beyond the fuller return model already implied by the asset's history?”
Consider an invented example. An asset earns 14%, and an exposure-based normal-return reference accounts for 10%. The remaining outcome is 4 percentage points. If a fuller historical model, including its fitted baseline, accounts for 13%, the remaining surprise is 1 point. Both comparisons use the same realized return but measure different remainders. These are educational figures, separate from the research results.
The intercept is not simply the stock's raw historical average return. It is a fitted baseline after the historical regression accounts for the included factor contributions. In the implementation, the primary normal-return model is estimated in excess-of-cash terms and excludes the fitted intercept; the surprise model is estimated on raw returns and includes it. The fits and cash treatment differ, so this comparison does not isolate an otherwise identical model with only one constant changed. Both models use prior estimation data and compound their normal daily returns across the evaluation window.
On the matched one-year sample, the average Alpha Score slope is +5.96 percentage points for absolute returns, +3.62 after market-risk adjustment and +4.56 after multifactor-risk adjustment. The multifactor-adjusted 95% interval is +2.94 to +6.18 points. In contrast, the stricter multifactor-model surprise estimate is -1.18 points, with an interval spanning zero. The study supports a positive exposure-adjusted association under the specified models, rather than prediction of surprises beyond the fitted historical model.
An interval spanning zero means these observations do not establish a statistically clear additional relationship on that surprise target. It does not prove that every possible additional relationship is exactly zero. One possible explanation is that the score organizes persistent differences already visible in historical return behavior. This is an interpretation, not proof of which input creates the association. It supports a careful distinction between an informative ranking and demonstrated incremental information beyond the fuller historical reference.
The comparison uses the same eligible observations for each primary return basis. Some historical factor-return inputs are unavailable, so the matched comparison has fewer observations than the original full absolute-return study. The interactive summary reports the count for the selected horizon. Alpha Score is the name of the ranking metric; model-adjusted return statistics are separate research outcomes.
How to read the research graphs
The study returns to a simple question: when the score was higher, what happened to the asset's return afterward? It compares scores with completed adjusted-close returns over 30 calendar days, 90 calendar days and one calendar year.
The five-group chart divides assets into Alpha Score quintiles each month. Q1 contains the lowest-scoring fifth and Q5 the highest. It displays the selected return measure within each group, averaged across observation months. Rising group averages indicate that the ranking was associated with that outcome in this sample.
The regression chart estimates the average return difference associated with a one-standard-deviation higher Alpha Score. Its unit is percentage points of the specified holding-period return. It describes an average fitted relationship rather than a target return for an individual investment.
Both views use the same underlying research panel. The Alpha Score groups emphasize ranking; the regression uses the numerical score differences. Comparing them helps show whether a finding appears only in one summary or across different ways of looking at the relationship.
The horizon and return-basis selectors change the actual study results displayed in the graphs. The return basis can be absolute, benchmark-relative or exposure-adjusted abnormal. The controls do not refit the models, change factor settings or generate new research outcomes. The separate educational calculators use invented numbers and are labeled as illustrations.
Why individual forecasts and group comparisons differ
A useful ranking can coexist with substantial unexplained movement in individual stocks. Company announcements, changing expectations, liquidity and influences missing from the model can cause large deviations from a forecast. A favorable expected return does not prevent disappointing company-specific news.
In the additional historical forecast exercise, the simple score-based rule explained only a small part of the variation in individual relative risk-adjusted outcomes. Its usefulness therefore needs to be assessed against the question being asked: predicting a particular stock's exact return is different from comparing the average outcomes of higher- and lower-scoring groups.
How diversification changes forecast uncertainty
Some company-specific surprises can offset when investments are considered together. Diversification does not create information in a score; it can make existing information easier to observe in a group average. The benefit depends on how the investments and their forecast errors move together.
This illustration assumes equal uncertainty for every investment and the same degree of shared error for every pair. Its assumptions are invented for teaching and are separate from the measured research results. When errors move together, increasing the number of names leaves a floor under uncertainty. The practical question is how much independent information the positions contribute, rather than simply how many names the portfolio holds.
A chronological test of larger baskets
The historical analysis was extended to compare an equally weighted basket of higher-scoring stocks with an equally sized basket of lower-scoring stocks. The exercise ranged from concentrated baskets to diversified baskets and the complete highest- and lowest-scoring groups.
A separate forecasting rule translated differences in Alpha Score into anticipated differences in group returns. At each observation date, it was fitted only to earlier holding-period outcomes that had already finished, with at least 36 completed monthly observations required before a forecast. This distinguishes the score used to rank assets from the additional rule used to forecast their return difference.
For each eligible month, baskets were repeatedly drawn from the existing high- and low-score groups using the same selection method at every breadth. Repeated draws show the influence of which companies were selected; they do not create additional independent years of market history.
How forecast error changed with breadth
For the one-year multifactor-risk-adjusted group difference, forecast error fell from 50.4 percentage points for concentrated baskets to 10.8 points for diversified baskets, and 10.1 points for the complete score groups.
The measure is root mean squared error (RMSE): a summary of the size of prediction misses that gives more weight to larger misses. It concerns the difference between two groups, rather than the forecast error of one investment or the volatility of a managed fund. The graph's stages describe increasing breadth without disclosing exact stock counts.
The remaining error is material. Broader baskets reduced sensitivity to individual selections, while shared changes across the group and changes in the score/return relationship could still affect outcomes.
Comparing anticipated and observed differences
For the complete highest- and lowest-score groups, the average anticipated one-year multifactor-risk-adjusted difference was 8.9 percentage points. The average observed difference was 12.5 points across 127 eligible monthly comparison windows.
These are averages across evaluation windows. They do not establish that each forecast was close to its eventual outcome. They show why an informative ranking can have economic relevance at the group level while leaving substantial uncertainty in individual returns.
What this exercise can establish
Chronological fitting uses only previously completed outcomes, but the current universe and scoring specification were selected as of the analysis date. Input availability, selection and survivorship effects still matter. The original score study, matched risk-adjusted comparison and later basket exercise have different eligible samples and should be read as separate tests.
Holding periods overlap, and repeated baskets reuse the same months. The figures describe higher-score-minus-lower-score holding-period outcomes before fees, trading, financing and other implementation costs. They are research evidence about ranking and aggregation, rather than a managed-portfolio track record. Diversification leaves shared risk, model error and the possibility of loss in place.
Bringing the investment process together
We assess relative expected opportunity, tilt the portfolio toward the stronger forecasts within its constraints, and use factor and co-movement estimates to manage the combination of positions. A stock can contribute through its expected return, through diversification, or through both.
The research is useful when it informs those decisions. It documents economically meaningful historical ranking differences under the stated return definitions. Portfolio accuracy, realized implementation costs and managed investment performance require their own direct evaluation.
Why portfolio construction matters
Different stories can conceal the same bet
An innovation narrative, a cash-flow narrative and a competitive-advantage narrative can lead to holdings that respond similarly to the same market forces. Different labels do not establish independent return sources.
In an anonymized example of three real hedge funds, the strategies presented with different investment narratives had pairwise return correlations between 0.83 and 0.90. The lesson is to examine their common behavior. Correlation flags overlap; identifying the particular exposures behind it requires further analysis. The example illustrates portfolio construction and is separate from the LRT Alpha Score study.
If several positions load on the same underlying influence, an adverse move can damage them together. Treating each position as an independent opportunity can consume the same risk budget repeatedly. Portfolio construction asks what the combined exposures are before deciding how much capital to allocate.
Why use quarter Kelly?
The Kelly criterion links position size to the estimated advantage and payoff distribution. Full Kelly maximizes expected long-run logarithmic wealth growth when those assumptions are correct. See Kelly's original paper.
In the simplified even-money wager, a 55% chance of winning and a 45% chance of losing imply a full-Kelly stake of 10% of current capital. Our example uses one-quarter Kelly: 2.5% of current capital. This is 25% of the calculated 10% stake.
Why use a smaller fraction? The estimated edge can be too optimistic, payoffs can change and relationships can weaken. Quarter Kelly reduces how much capital depends on getting those estimates right. It deliberately gives up some modeled growth to leave a margin for estimation error and changing conditions. In this wager, cutting the stake to one-quarter also cuts one-round return volatility to one-quarter.
Consider an invented sensitivity check: the win probability turns out to be 52% rather than the estimated 55%. The full-Kelly stake for that weaker edge would be 4%. A 10% stake is then too aggressive and produces negative expected logarithmic growth; the 2.5% quarter-Kelly stake retains positive expected logarithmic growth in that scenario. The graph shows both assumptions, making the reason for fractional sizing visible.
The trade-off remains explicit. Under the original 55% assumption, modeled geometric growth is approximately 0.22% per round at quarter Kelly versus 0.50% at full Kelly. These are mathematical outcomes of the stated repeated-wager model. Fractional sizing reduces exposure; it does not turn a nonexistent edge into a profitable one or establish a drawdown limit. Research on risk-constrained Kelly allocation examines drawdown control separately from the growth objective.
Real equities have continuous, changing and correlated payoffs. Their weights require joint portfolio analysis, liquidity and risk constraints. Quarter Kelly is the sizing illustration here, rather than a claim that LRT's production optimizer implements a particular Kelly fraction.
Correlation still matters. In the invented wager, three labels on exactly the same outcome, each sized at 2.5%, create a combined 7.5% exposure to one outcome. The smaller individual stakes do not make those wagers independent. The risk budget must be assessed for the combination.
Why severe losses change the compounding problem
A 50% loss requires a 100% gain to recover. A 70% loss requires approximately 233%. The arithmetic follows from the smaller amount of capital left after the loss. Oversized positions, common factor exposure and financing pressure can make recovery much harder even when the original investment ideas were sensible.
Borrowing and gross exposure also matter in a hedged portfolio. A small net dollar exposure does not by itself imply small factor risk, small gross positions or manageable funding demands. Construction considers those dimensions together, including what a shared adverse move would do to the portfolio and its ability to maintain positions.
This connects the risk work to the investment objective. A risk model helps allocate capital among opportunities while keeping unwanted exposures from dominating the result. The discussion in LRT's paper on betting discipline and portfolio construction develops this connection between the investment edge and the size of the risk taken.
Reduce market exposure through index hedges
Index hedges provide a way to reduce the portfolio's broad market exposure while maintaining the selected stock positions. The hedge decision considers the market sensitivity of the holdings and the exposure the portfolio is intended to retain.
Broad market exposure and other factor concentrations are separate parts of the construction problem. An index hedge can reduce a common market component while leaving other shared sensitivities among the stocks. Position selection, sizing and exposure constraints address those additional concentrations. Hedge sizing also requires review as holdings, prices and estimated sensitivities change.
How to read the uncertainty bars and period checks
Research charts show 95% confidence intervals around the estimated historical means and score coefficients. The high-minus-low interval is calculated from the chronological monthly group differences directly. It is not the difference between the confidence limits of the two group means. The HAC calculations allow for the dependence introduced by overlapping forward-return windows.
The separate robustness figure examines the absolute-return score relationship across earlier, later and recent samples and alternative estimation treatments. Estimates were positive in each displayed period at each horizon, and the displayed intervals were above zero. Recent is a subset of later; these are sensitivity checks within the historical sample, not independent holdout tests. They use the full absolute-return sample, rather than the smaller matched exposure-adjusted sample. A confidence interval for an average relationship is distinct from an individual-stock prediction interval.
A short glossary
Descriptor: A measurement of an asset characteristic at a particular date.
Factor return: A return series representing a common influence or a defined basket comparison.
Exposure or loading: An estimated sensitivity of an asset's returns to a factor series, conditional on the fitted model.
Residual: The portion of an observed return left after subtracting the fitted baseline and factor contributions.
Regularization: A fitting technique that penalizes coefficient size to help control unstable or overly complex estimates.
Standardization: Expressing a measurement relative to a comparison group's mean and dispersion.
LRT Alpha Score: LRT Capital's proprietary combined, signed ranking metric for assessing relative asset attractiveness within a specified universe.
LRT Alpha Framework: The connected analytical process surrounding the score, spanning factor engineering, joint estimation, residual analysis, scoring, portfolio construction and research validation.
Quintile: One of five groups formed by sorting observations. Each group contains approximately one fifth of the available assets.
Percentage point: The arithmetic difference between two percentages. A return of 5% compared with 3% is a difference of 2 percentage points.
Abnormal return: The difference between an observed return and the normal return specified by a stated comparison model.
Historical-model surprise: An abnormal return measured against a fuller normal-return model that also includes the fitted historical intercept.
Expected return: An estimate made before the outcome is known, describing a central expectation across possible outcomes over a stated horizon.
Basket: A group of investments considered together. The chronological forecast exercise uses equal weights within each comparison group.
Root mean squared error (RMSE): A measure of forecast misses calculated by squaring the errors, averaging them and taking the square root. Larger misses receive more weight, and the result uses the same units as the forecast target.
Further reading
LRT's investment process is a very bright white box: we want investors to examine the evidence and understand how it informs our decisions. This guide is one part of a broader library of data and research we can share. Start with the LRT factor explorer to compare historical factor returns, Sharpe ratios and correlations, inspect definitions and current compositions, and download the data. Ask LRT Capital about additional research and supporting materials for a deeper review of the investment process.
For a classic example of constructing return factors from sorted portfolios, see the Kenneth French Data Library's construction notes. For an introduction to joint linear estimation and regularization, see the scikit-learn linear-model guide. These are general educational references, separate from LRT's proprietary factor definitions and the source of this study's results.