Is Quant Investing Just Factor and Risk Models All the Way Down?
Oct 2019-Sep 2026: six factors explained 93% of AVUV's tracking-error variance vs VTI but 30% of QUAL's. What quant risk models do, and where they fail.
Read a quant fund’s prospectus and you find signals, models and optimizers. Under those words sit two pieces of math: a model of which stocks should earn more, and a model of how stocks move together. The claim that systematic investing is “factor and risk models all the way down” borrows Corey Hoffstein’s line about long/short portfolios, which our guide to why every portfolio is long/short traces. This guide tests the claim one layer at a time, with our own measurements of seven funds a DIY investor can buy.
For the systematic funds most DIY investors own, the claim holds up well. MSCI publishes the recipe for its multiple-factor indexes, and it is a factor score, a Barra risk model and an optimizer with constraints.1 Over October 2019 to September 2026, six academic factors accounted for 93% of the variance in AVUV’s return difference from VTI.2 The claim breaks in three places: it cannot say why a factor pays, linear factor models miss option-like strategies such as trend following, and high-frequency market making runs on speed and inventory. One result points the other way for fund buyers: over the same window the same six factors accounted for only 30% of QUAL’s difference from VTI, because QUAL’s index ranks stocks on MSCI’s own quality measures, which the Fama-French factors capture only in part.
The short version
- Quant funds run two models. A return model ranks stocks; a risk model estimates how they move together. A risk factor needs no premium: industries and countries are in risk models because they cause losses.
- Risk models exist because 3,000 stocks have about 4.5 million distinct variances and covariances, and ten years of monthly data cannot estimate them. A factor structure cuts the count to a few thousand.
- Over Oct 2019 to Sep 2026, six Fama-French factors accounted for 90% to 95% of the tracking-error variance of three small-value funds against VTI, and 30% to 66% for three MSCI factor-index funds. Adding industry returns raised USMV’s adjusted figure from 52% to 87%.
- Tracking error moves. For a 30% small-value sleeve against the U.S. market, realized 12-month tracking error ranged from 1.1% to 6.2% across windows ending 1999 to 2026.
- Size a tilt by the tracking error you can sit through, check which model built a fund before reading its regression, and expect two managers with the same factor bet to move together.
Two models with different jobs
A quant manager splits the problem in two. The return model, often called the alpha model, ranks securities by expected return using signals such as valuation, profitability and past returns. The risk model answers a different question: how much could this portfolio, or its difference from a benchmark, move in a year, and which exposures would cause it?
| Return (alpha) model | Risk model | |
|---|---|---|
| Question | Which stocks will earn more? | What could make this portfolio lose money, and how much? |
| Output | Scores or expected returns | A covariance matrix: factor covariances plus stock-specific variances |
| Must its factors earn a premium? | Yes, that is the point | No. Industries, countries, currencies and market beta all qualify |
| How it is judged | Out-of-sample returns | Forecast volatility against realized volatility |
MSCI’s Barra U.S. equity model (USE4) shows the difference. It has 12 style factors, and beside them a country factor and 60 industry factors based on GICS.3 A semiconductor industry factor is in the model because a portfolio overweight in semiconductors loses money when that industry falls, whatever the portfolio’s value or momentum scores say. The model assumes no premium for it.
Whether a factor that does earn a premium is paid for bearing risk, for exploiting other investors’ mistakes or for working around trading frictions is a separate argument, which our guide to models beyond Fama-French covers. A risk model works the same under any of those explanations. Kozak, Nagel and Santosh showed why the explanations are so hard to separate: if there are no near-arbitrage opportunities, the stochastic discount factor can be written as a function of a few dominant sources of return variation, and this holds “even in an economy in which all cross-sectional variation in expected returns is caused by sentiment.”4 A sizable premium has to line up with the main ways stocks move together, or it would be close to free money, whatever causes it. That result is the best theoretical support for the all-the-way-down claim, and it is silent on why the factor pays.
Why every quant fund runs a risk model
Start with a universe of 3,000 stocks. Its covariance matrix has 3,000 variances and 4,498,500 distinct covariances, 4,501,500 numbers in all. Ten years of monthly returns give 120 observations per stock. A sample covariance matrix built from 120 observations has rank of at most 119, so at least 2,881 of its 3,000 dimensions show zero estimated variance. An optimizer handed that matrix will find portfolios that look riskless in the sample and are not. Our test of mean-variance optimization out of sample shows how far estimation error moves optimized weights even with ten assets.
A factor model fixes the counting problem by assuming each stock’s return is a set of exposures times a few factor returns, plus a stock-specific piece:
Here holds each stock’s exposures to factors, is the covariance of the factor returns, and is a diagonal matrix of stock-specific variances. In a fundamental model like Barra’s, the exposures are observed: industry membership and standardized characteristics such as size or book-to-price. Each period a cross-sectional regression of stock returns on those exposures gives that period’s factor returns, and their history gives .5 With USE4’s 73 factors, the model needs factor covariances plus 3,000 specific variances, about 5,700 numbers instead of 4.5 million.
Engineers will recognize the structure: a low-rank matrix plus a diagonal, the same shape as factor analysis in statistics. The price is the diagonal. It assumes stock-specific returns are uncorrelated with each other. When a group of stocks shares an exposure the model does not have, their “specific” returns move together and the model understates the risk of owning them together. MSCI’s USE4 factsheet lists two corrections that address related problems, an optimization bias adjustment and a volatility regime adjustment.3 The first exists because optimizers drift toward portfolios whose risk the model underestimates; the second because volatility changes faster than a long estimation window can follow.
What the risk model decides
Inside a quant process the risk model does three jobs. It strips out bets nobody chose, such as the industry tilt or market-beta tilt a value screen produces by accident. It sets a tracking-error budget, the expected volatility of the portfolio’s return minus its benchmark’s. And it sizes positions so that no single stock’s specific risk dominates the result.
The budget matters because of what Richard Grinold called the fundamental law of active management. The information ratio, a strategy’s expected active return divided by its tracking error, is roughly the skill of each forecast times the square root of the number of independent forecasts a year.6 Clarke, de Silva and Thorley added a third term, the transfer coefficient, which measures how closely the weights a manager can hold track the weights the forecasts call for:
A signal with an information coefficient of 0.05 applied to 400 independent bets a year implies an information ratio near 1.0 with no constraints. Clarke, de Silva and Thorley reported that “in practice, managers often work under constraints that produce transfer coefficients ranging from 0.3 to 0.8.”7 Ruling out short sales is one of the constraints that lowers it. In a cap-weighted benchmark most stocks already have small weights, so the most a long-only fund can underweight a stock it expects to lag is that small weight. A long-only factor fund therefore passes through only part of its signal, and its loading on a long-short research factor stays well below what an unconstrained portfolio could reach. Our guide on factor loadings and expected returns explains what a loading of 0.5 does and does not mean.
When the return model and the risk model disagree
An optimizer sees the world only through the risk model. If a return signal, say a quality score, is not one of the risk model’s factors, the optimizer treats bets on it as stock-specific risk that diversifies away, and it piles in. If the signal is one of the factors, the optimizer charges for it and shrinks the bet. Axioma researchers Anureet Saxena and Robert Stubbs called these factor alignment problems and described the cause as “the optimizer’s propensity to neutralize the exposure of the optimal portfolio to all systematic risk factors that are included in the risk model, and inadvertently load up on others which are omitted.”8 Their proposed fix adds the part of the return signal the risk model cannot explain as one extra factor, so the optimizer stops treating it as free. The published version, with Sebastián Ceria, appeared in 2012.9
MSCI’s own researchers reached a similar conclusion from the other side. Lee and Stefek found that “aligning risk factors with alpha factors may improve the information ratio of optimized portfolios, even if doing so lowers the overall accuracy of risk forecasts.”10 A risk model tuned only for forecasting accuracy can produce worse portfolios than one tuned to the signals it will be used with. The two models have to be built together.
For a fund buyer, the alignment problem shows up in reverse. When you test a fund with the Fama-French factors, you are using a risk model different from the one that built it. If the fund was built on other factors, such as an optimizer’s industry factors or an index provider’s own quality score, much of its active risk will land in your regression’s residual. The section after next measures how much.
Published recipes
MSCI documents its factor index construction in detail, which makes it the clearest public example of the full pipeline. The MSCI Core Multiple-Factor Indexes, in the April 2025 methodology, combine value, momentum and quality into one score and choose weights with the Barra Open Optimizer on MSCI’s long-horizon global equity model, maximizing the score minus a penalty for active risk.1 The constraints:
- Predicted tracking error at most 2%, and active specific risk at most 1%.
- Predicted beta between 0.98 and 1.02.
- Each GICS sector within 5 percentage points of its parent-index weight.
- One-way turnover at most 7.5% per quarterly rebalance.
- Target styles held between +0.1 and +0.6 standard deviations (−0.6 to −0.1 for earnings variability and leverage), and styles the index does not target within ±0.1.
The methodology also lists what to loosen when no portfolio satisfies every constraint: turnover in 1.5-point steps up to 15%, tracking error in 0.2-point steps up to 3%, and specific risk in 0.1-point steps up to 1.5%.1 The factor score says what to buy; the risk model and the constraints decide how much.
USMV, the iShares MSCI USA Minimum Volatility ETF, is the case where the risk model is the whole strategy. MSCI uses the Barra Open Optimizer with a Barra equity model to find the lowest-predicted-risk portfolio, with stock weights capped at the lower of 1.5% or 20 times the parent weight, sectors within 5 points of the parent, and one-way turnover at most 5% per quarterly review.11 Our minimum-volatility guide covers its live record.
Avantis and Dimensional describe their small-value funds differently. Avantis says it uses “information in current market prices and company financials to identify differences in expected returns among securities, seeking to overweight securities with higher expected returns.”12 Dimensional’s summary prospectus for its U.S. small-cap value ETF describes “an integrated investment approach that combines research, portfolio design, portfolio management, and trading functions,” and says the adviser may adjust holdings “based on shorter-term considerations, such as a company’s price momentum, short-run reversals, and investment characteristics.”13 Neither document states a numeric tracking-error budget the way MSCI’s methodologies do.
How much of a factor fund’s difference from VTI is factors?
To test the claim on funds people own, we regressed each fund’s monthly return minus VTI’s on six factors, the Fama-French five plus momentum, over October 2019 to September 2026. That is 84 months, and the start is set by AVUV’s history.2 In-sample, ordinary least squares splits the variance of that return difference exactly into a factor part and a residual part:
is the fund’s loading on each factor relative to VTI, the factor covariance over the same months, and the residual variance. The factor share in the table is the regression’s R².
| Fund | Tracking error vs VTI | Six factors’ share | Residual TE | Largest contributions, % of variance |
|---|---|---|---|---|
| VBR Vanguard small-cap value index | 10.50% | 90.3% | 3.27% | Size 51, value 38 |
| AVUV Avantis small-cap value | 15.23% | 92.6% | 4.15% | Size 53, value 38 |
| DFSVX Dimensional small-cap value | 13.66% | 95.0% | 3.06% | Size 53, value 40 |
| USMV iShares MSCI USA min vol | 8.78% | 55.3% | 5.87% | Market 35, profitability 10 |
| QUAL iShares MSCI USA quality | 3.79% | 29.7% | 3.18% | Profitability 14, size 12 |
| MTUM iShares MSCI USA momentum | 10.95% | 66.1% | 6.37% | Momentum 52, profitability 13 |
Summitward calculation. Monthly return minus VTI’s regressed on Fama-French five factors plus momentum, October 2019 to September 2026 (84 months). Tracking error is annualized. Contributions are Euler shares of tracking-error variance. Data: Ken French library (202609 CRSP database), Yahoo Finance adjusted closes. Historical measurement over the stated window.
For the three small-value funds the claim holds almost exactly. Six factors account for 90% to 95% of the variance in their return difference from VTI, and size and value supply nearly all of it. Their Fama-French regressions describe what you own. Over each fund’s longer history against VTI the shares are similar: 88% for VBR since February 2004 and 92.5% for DFSVX since July 2001.
The three MSCI index funds look different, and only one of them uses an optimizer. QUAL’s index scores stocks on return on equity, debt to equity and earnings variability within each GICS sector, weights them by score times market cap, and sets each sector to its parent weight.14 MTUM’s index ranks stocks on 6- and 12-month price momentum, skipping the latest month and scaled by each stock’s volatility, and also weights by score times market cap.15 USMV’s largest source of tracking error was the market factor itself: its loading on the market relative to VTI was −0.28, so it lags in rallies and cushions declines. Even so, 45% of its tracking-error variance sits in the residual. QUAL’s residual is 70%, and MTUM’s is 34%.
Commercial risk models such as Barra’s have industry factors and the Fama-French model does not, so we added the returns of 11 industries minus the market, from Ken French’s 12-industry portfolios.16 With 17 regressors on 84 months, raw R² rises mechanically, so the comparison uses adjusted R²:
| Fund | Six factors, adjusted R² | Plus 11 industries, adjusted R² |
|---|---|---|
| VBR | 89.6% | 94.5% |
| AVUV | 92.0% | 95.2% |
| DFSVX | 94.6% | 95.6% |
| USMV | 51.9% | 86.5% |
| QUAL | 24.2% | 51.3% |
| MTUM | 63.5% | 68.8% |
Summitward calculation, same window and regressions as above, with value-weighted French 12-industry returns minus the market return (“Other” omitted). Adjusted R² corrects for the number of regressors.
Industry tilts explain most of what the six factors miss in USMV: adjusted R² rises from 52% to 87%. MSCI caps USMV’s sector weights within 5 points of the parent, and that still leaves room for industry bets large enough to account for about a third of its tracking-error variance. Over USMV’s own history since November 2011 the same comparison runs from 56% to 86%. QUAL improves from 24% to 51% and still leaves about half unexplained. That remainder bundles stock-specific bets with the difference between MSCI’s quality score, which rewards high return on equity, low debt and stable earnings, and the Fama-French profitability factor, and this data cannot separate the two.
So the answer depends on whose factors you use. Each of these funds is built from factor scores, and USMV from a risk model as well, but the Fama-French lens sees the small-value funds almost completely and the MSCI funds only in part.
Exposure and risk contribution are different numbers
A portfolio’s active loadings are the weighted sum of its funds’ loadings, and each factor’s contribution to tracking error depends on the loading, the factor’s volatility and its correlation with every other bet in the portfolio:
Take 70% VTI and 30% AVUV over the same 84 months. Its tracking error against VTI was 4.57%. Its active loadings were +0.27 on size, +0.16 on value and +0.06 on profitability. Size contributed 53% of the tracking-error variance, value 38%, and profitability 1.2%. The profitability loading was about 40% of the value loading, but its contribution was about one-thirtieth as large: variance scales with the square of the loading, profitability’s volatility in the window was 10.5% a year against value’s 14.7%, and its correlation with the size factor was −0.22, which offset part of the size bet.2
The window matters too. Size and value returns had a correlation of +0.39 over October 2019 to September 2026, against +0.02 since 1963, so the two bets reinforced each other. Swap in the factor volatilities and correlations since 1963 and the same portfolio’s estimated tracking error falls from 4.57% to 3.43%.
What you pair a tilt with matters. Splitting the 30% evenly between AVUV and DFSVX cut tracking error only from 4.57% to 4.31%, because both funds carry the same size and value bets. MTUM on its own strayed from VTI almost as much as DFSVX did (10.95% against 13.66%), yet splitting the 30% between AVUV and MTUM cut tracking error to 2.42%, because momentum’s bets differ from small value’s. Splitting it with QUAL gave 2.19%, though mostly because QUAL’s own bets are small. Try your own mix:
A tilt’s tracking error moves around
Any tracking-error figure, from a risk model or a regression, is an estimate for a window. To see how much it moves, we took a 30% DFSVX sleeve against the U.S. market (Ken French’s market return) from March 1993 to September 2026. Its tracking error over the whole period was 3.36%. Each month from February 1998 to September 2025, we compared the trailing 60-month tracking error with the tracking error the next 12 months produced.2
Realized 12-month tracking error ranged from 1.09%, in the 12 months to March 2013, to 6.22%, in the 12 months to February 2021. The median ratio of realized to forecast was 0.98, so the trailing number was about right at the median, but the 10th and 90th percentiles of the ratio were 0.53 and 1.70.
Some of that spread is sampling noise: a 12-month standard deviation is a rough estimate even when the true value never changes. Under i.i.d. normal returns, the ratio would fall outside 0.64 to 1.34 in 10% of windows from noise alone. It fell outside that band in 35% of months, 19% above and 16% below. The windows overlap, so those shares describe history and do not count independent trials. The largest misses came from forecasts made in late 1998 and early 1999, when trailing tracking error near 2.4% was followed by about 5% in the year of the dot-com peak; from February 2020, when 3.0% was followed by 6.2%; and from October 2016, when 2.0% was followed by 4.3%.
For a DIY investor the practical reading is to plan for a tilt’s tracking error to run well above its trailing figure in a bad year; in this sample it reached about twice the forecast. A risk model cannot remove that uncertainty. It can tell you which bet the tracking error comes from.
Crowding: shared models, shared exits
If many managers rank stocks on similar signals and control risk with similar models, they hold similar portfolios, and they can be forced to sell them at the same time. Looking at August 2007, Amir Khandani and Andrew Lo found that a simple contrarian long/short equity strategy lost 1.16%, 2.83% and 2.86% on August 7, 8 and 9, a three-day loss of 6.85%, then gained 5.92% on August 10. Levered 8 to 1, the same strategy would have lost a little over a quarter of its assets in three days. They attributed the losses to the rapid unwinding of one or more large quantitative market-neutral portfolios, and wrote that “any explicit factors used to construct that portfolio would have generated a loss for other portfolios with the same factor exposures.”17
Lasse Pedersen estimated that a value-plus-momentum strategy scaled to 6% annual volatility lost about 25% in four days that week, and named a limit every risk model shares: the risk model “does not capture liquidity risk.”18 Covariances estimated from normal months say nothing about what happens when everyone holding the same factor portfolio sells it at once. Lou and Polk later showed the same effect in momentum: when momentum stocks become unusually correlated with each other, a sign of crowded trading, their returns subsequently reverse.19
The same logic applies to unlevered retail factor funds. AVUV and DFSVX come from different managers with different processes, yet over October 2019 to September 2026 the parts of their returns the six factors could not explain had a correlation of 0.71.2 Owning both may be worth it for other reasons, such as tax lots or manager risk, but it does little for tracking error.
Where the factor lens stops working
Some quant strategies fit the factor framework so closely that they trade its leftovers. Others fit it badly.
| Strategy | Fits a linear factor model? | Evidence |
|---|---|---|
| Statistical arbitrage | Yes. Its signal is the factor model’s residual | Avellaneda and Lee traded mean reversion in residuals from PCA and sector-ETF factor models: Sharpe 1.44 for the PCA version over 1997–2007, falling to 0.9 over 2003–2007, after costs.20 |
| Machine-learning stock selection | Partly. Nonlinear, but built on familiar characteristics | Gu, Kelly and Xiu found trees and neural networks predicted best, and every method agreed the dominant signals were momentum, liquidity and volatility.21 What that means for a small-value fund is in our guide to building a better one. |
| Trend following (managed futures) | No. Payoffs are option-like | For trend-following funds, eight asset-class indexes explained an adjusted R² of 1.0%; Fung and Hsieh’s lookback straddle factors explained 47.9%.22 |
| Merger arbitrage | No. Beta changes with the market | Mitchell and Pulvino found market beta near zero in flat or rising markets and about 0.5 in months the market fell 4% or more, like writing index puts.23 |
| High-frequency market making | No. Returns come from speed and inventory | Baron, Brogaard, Hagströmer and Kirilenko found that differences in relative latency account for large differences in high-frequency traders’ performance; in their working paper the median firm’s four-factor alpha was about equal to its raw return.24 |
Statistical arbitrage supports the thesis in an unexpected way: the factor model defines the trade even when the profit comes from what the model leaves out. Fung and Hsieh put the failure case bluntly: an investor using a linear factor model on standard benchmarks “would have concluded that trend followers had no systematic risk.”22 Agarwal and Naik found that many equity hedge-fund strategies have payoffs resembling a short put on the market, a left-tail risk that mean-variance analysis ignores.25 If you own a managed-futures fund, an options income fund or a merger-arbitrage fund, a Fama-French regression will misdescribe how it behaves in a crash: a straddle-like payoff tends to gain in large moves, and a short-put payoff loses. Our guide to managed futures covers that case.
What I recommend
- Hold a total-market fund and you can skip all of this. VTI has no tracking error against itself, and a global index fund has none against its own benchmark. Risk models matter once you deviate.
- Set a tracking-error budget before you add a tilt. A 30% AVUV sleeve carried 4.6% tracking error against VTI over October 2019 to September 2026, so a two-standard-deviation year meant trailing VTI by about 9 points, and the stability test shows the figure can run higher in a bad year. If that gap would make you sell, size the tilt smaller. Pair this with the break-even premium a tilt needs after fees and taxes.
- Check which model built a fund before trusting its regression. For index and rules-based small-value funds, a Fama-French regression captured roughly 90% or more of what made them differ from the market. For MSCI factor funds such as USMV, QUAL and MTUM, read the index methodology and look at the industry weights as well.
- Know what you pay for in the residual. Residual tracking error is the part of a fund no factor exposure reproduces: its industry tilts, its stock picks, its trading. It can help or hurt. A fund charging more than a plain factor fund should be able to say what its residual is for.
- Diversify across different bets. Two small-value funds from different managers moved together even after removing the factors. A second fund with a different bet, such as quality or momentum, changes tracking error far more than a second manager with the same bet.
Building your own stock-level risk model is not worth it for a portfolio of a few funds. The fund-level version in the calculator above, or Summitward’s portfolio factor analysis, answers the questions a DIY investor faces: which factors you own, and how far that can put you from the market.
Key Takeaways
- Quant investing runs on two models: a return model that ranks stocks and a risk model that estimates how they move together. Risk factors such as industries need no premium.
- A factor structure turns the 4.5 million covariances of a 3,000-stock universe into a few thousand estimable numbers, at the cost of assuming stock-specific returns are uncorrelated.
- Over October 2019 to September 2026, six Fama-French factors explained 90% to 95% of three small-value funds’ tracking-error variance against VTI, and 30% to 66% for three MSCI factor-index funds, where industry tilts filled much of the gap for USMV.
- A 30% small-value sleeve’s realized 12-month tracking error ranged from 1.1% to 6.2% across windows ending 1999 to 2026, far more spread than sampling noise alone produces.
- The factor lens fails for option-like strategies such as trend following and merger arbitrage, and for high-frequency market making.
Frequently Asked Questions
What is a risk model in investing?
A model that estimates how volatile a portfolio is and which exposures cause it. Most commercial equity risk models write each stock’s return as exposures to a few dozen factors, such as industries, countries and styles, plus a stock-specific part, and estimate the covariance of those factors from history.
What is the difference between a factor model and a risk model?
“Factor model” usually means a model of expected returns, like the Fama-French five-factor model, whose factors are meant to earn premiums. A risk model uses factors to estimate covariance, and many of its factors, such as industries, are there because they cause losses; the model assumes no premium for them.
How much tracking error does a small-cap value tilt add?
It depends on the size of the tilt and the window. A portfolio of 70% VTI and 30% AVUV had 4.57% annual tracking error against VTI over October 2019 to September 2026, or 3.43% using factor volatilities since 1963. A 30% DFSVX sleeve against the market averaged 3.36% over 1993 to 2026 and ranged from 1.1% to 6.2% across 12-month windows.
Do quant funds all use the same risk model?
No, but several commercial vendors sell risk models, and published index methodologies name them: MSCI builds its minimum-volatility and multiple-factor indexes with Barra models. Khandani and Lo argued that the August 2007 losses implied a growing common component across quantitative equity funds.
Is all quant investing factor investing?
Most long-only systematic equity funds are well described by factors. Statistical arbitrage trades the residuals of factor models. Trend following, merger arbitrage, options strategies and high-frequency market making have option-like or speed-driven payoffs that linear factor models describe poorly.
What is the transfer coefficient?
The correlation between the weights a manager actually holds and the weights its forecasts call for. Constraints such as no short selling and turnover limits lower it; Clarke, de Silva and Thorley reported values from 0.3 to 0.8 for constrained managers. The expected information ratio scales with it.
Related Guides
- Every Portfolio Is Long/Short Active weights, Active Share and tracking error.
- Beyond Fama-French q-factors, AQR’s factors, and why alpha depends on the model.
- Do Higher Factor Loadings Mean Higher Expected Returns? The return side of the same loading gap.
- Mean-Variance Optimization Out of Sample Why optimizers amplify estimation error.
- Minimum Volatility and Betting Against Beta USMV, the fund built by a risk model.
- Fama-French Factor Analysis How to read loadings, t-statistics and R².
- Is Factor Investing Dead? Replication, decay, costs, and a tilt break-even calculator.
- The Engineer’s Guide to Systematic Investing Process over prediction.
Sources
- MSCI, “MSCI Core Multiple-Factor Indexes Methodology,” April 2025, sections 2.5 and Appendix III. msci.com
- Summitward calculations: monthly fund returns minus VTI’s regressed on Fama-French five factors plus momentum, October 2019 to September 2026 (84 months) and over each fund’s own history; exact in-sample split of tracking-error variance; industry check with French 12-industry portfolios; five fund mixes; and a 30% DFSVX sleeve against the French market return, March 1993 to September 2026, with trailing 60-month and next-12-month tracking error. Data: Ken French data library (202609 CRSP database), Yahoo Finance adjusted closes. Script and output: summitward-research
- MSCI, “Barra US Equity Model (USE4)” factsheet, 2013. msci.com
- Serhiy Kozak, Stefan Nagel and Shrihari Santosh, “Interpreting Factor Models,” Journal of Finance 73(3), 2018, 1183–1223. doi.org
- Jose Menchero, D.J. Orr and Jun Wang, “The Barra US Equity Model (USE4): Methodology Notes,” MSCI, August 2011, equations 1.5 and 3.1. copy hosted by top1000funds.com
- Richard C. Grinold, “The Fundamental Law of Active Management,” Journal of Portfolio Management 15(3), 1989, 30–37. doi.org
- Roger Clarke, Harindra de Silva and Steven Thorley, “Portfolio Constraints and the Fundamental Law of Active Management,” Financial Analysts Journal 58(5), 2002, 48–66. doi.org
- Anureet Saxena and Robert A. Stubbs, “Finance, Statistics, Accounting, Optimization and Some Alignment Problems,” Axioma Research Paper No. 28, November 2010. simcorp.com
- Sebastián Ceria, Anureet Saxena and Robert A. Stubbs, “Factor Alignment Problems and Quantitative Portfolio Management,” Journal of Portfolio Management 38(2), 2012, 29–43. doi.org
- Jyh-Huei Lee and Dan Stefek, “Do Risk Factors Eat Alphas?” Journal of Portfolio Management 34(4), 2008, 12–25. Quotation from MSCI’s abstract. doi.org
- MSCI, “MSCI Minimum Volatility Indexes Methodology,” July 2025, section 3.2. msci.com
- Avantis Investors, Avantis U.S. Small Cap Value ETF (AVUV) fund page, read November 2026. avantisinvestors.com
- Dimensional ETF Trust, Dimensional US Small Cap Value ETF summary prospectus, February 28, 2026 (SEC Form 497K). The ETF merged into the U.S. Small Cap Value Portfolio (DFSVX) as a share class on September 25, 2026. sec.gov
- MSCI, “MSCI Quality Indexes Methodology,” May 2025, sections 2.2 to 2.4 and Appendix VI (sector neutral quality). msci.com
- MSCI, “MSCI Momentum Indexes Methodology,” July 2025, section 2.2. MTUM tracks the MSCI USA Momentum SR Variant Index, which follows the MSCI USA Momentum Index with rebalance trades spread over several days. msci.com
- Kenneth R. French, Data Library: Fama/French 5 factors (2x3), momentum factor and 12 industry portfolios. Since the January 2025 release, U.S. returns are built from CRSP’s CIZ files. tuck.dartmouth.edu
- Amir E. Khandani and Andrew W. Lo, “What Happened to the Quants in August 2007?” Journal of Investment Management 5(4), 2007, 29–78, sections 4, 6 and 7 (working paper); and “What Happened to the Quants in August 2007? Evidence from Factors and Transactions Data,” Journal of Financial Markets 14(1), 2011, 1–46. doi.org
- Lasse Heje Pedersen, “When Everyone Runs for the Exit,” International Journal of Central Banking 5(4), 2009, 177–199. ijcb.org
- Dong Lou and Christopher Polk, “Comomentum: Inferring Arbitrage Activity from Return Correlations,” Review of Financial Studies 35(7), 2022, 3272–3302. doi.org
- Marco Avellaneda and Jeong-Hyun Lee, “Statistical Arbitrage in the US Equities Market,” Quantitative Finance 10(7), 2010, 761–782, Table 6. doi.org
- Shihao Gu, Bryan Kelly and Dacheng Xiu, “Empirical Asset Pricing via Machine Learning,” Review of Financial Studies 33(5), 2020, 2223–2273. doi.org
- William Fung and David A. Hsieh, “The Risk in Hedge Fund Strategies: Theory and Evidence from Trend Followers,” Review of Financial Studies 14(2), 2001, 313–341, Table 2. doi.org
- Mark Mitchell and Todd Pulvino, “Characteristics of Risk and Return in Risk Arbitrage,” Journal of Finance 56(6), 2001, 2135–2175. doi.org
- Matthew Baron, Jonathan Brogaard, Björn Hagströmer and Andrei Kirilenko, “Risk and Return in High-Frequency Trading,” Journal of Financial and Quantitative Analysis 54(3), 2019, 993–1024 (doi.org). Alpha comparison from the 2017 working paper, Swedish equities 2010–2014. cityu.edu.hk
- Vikas Agarwal and Narayan Y. Naik, “Risks and Portfolio Decisions Involving Hedge Funds,” Review of Financial Studies 17(1), 2004, 63–98. doi.org
Method
Fund returns are month-end changes in Yahoo Finance dividend-adjusted closes; factor and market returns are from Ken French’s library, files built from the 202609 CRSP database. Regressions are ordinary least squares with an intercept on the return difference between each fund and VTI. Every variance uses the n − 1 divisor, which makes the factor-plus-residual split exact within the window; the script asserts it. Mixes rebalance monthly and are gross of trading costs and taxes. DFSVX’s longer history against VTI starts in July 2001, VTI’s first full month. As a pipeline check, the script reproduces the AVUV loadings published in our factor-loadings guide to four decimal places. Code, data and printed output are in the summitward-research repository.
Author disclosure
Summitward has no business relationship with Avantis, Dimensional, iShares, MSCI, Vanguard or any firm mentioned here and receives no compensation from them. Figures labeled as Summitward calculations are historical measurements over the stated window, not forecasts. Nothing here is investment advice.
More in Investing & Portfolio
Browse all investing & portfolio guidesGet new guides by email
Evidence-based, no jargon. At most two emails a month. Unsubscribe any time.
Try it in Summitward
See portfolio factor analysis in action with your own financial data. Free to start, no credit card required.