Could Anyone Build a Better Small-Cap Value Fund Than AVUV or DFSV?
Five small-value funds with different designs, Oct 2019 to Sep 2026: none beat AVUV by more than luck explains. Could any fund, or machine learning, do better?
AVUV and DFSV are the two default ways to own U.S. small-cap value stocks in an ETF. Each charges about a quarter of a percent a year, holds between 790 and 1,044 stocks, and buys companies that are cheap relative to their book value and also profitable.1,8 Could someone build a fund that does better? And could machine learning or AI agents be how they do it?
My answer is that a better fund is possible on paper, but most of the upgrades people propose are already inside these two funds. The rest either take more risk or a different mix of it, which their factor loadings show, or would take decades of live returns to verify: at the 3.5% tracking error between AVUV and Dimensional’s small-value fund from 2019 to 2026, a fund that truly beat AVUV by 1% a year would need about 48 years of returns before a standard test had even odds of confirming it. Over the seven years they have all existed side by side, four funds built around different ideas trailed AVUV by 0.7 to 1.7 percentage points a year, and none of the gaps, in either direction, was large enough to rule out luck. That falls short of proof that AVUV and DFSV cannot be improved. Machine learning has a real research record, but most of its gains come from fast-moving signals in small, hard-to-trade stocks. When researchers make the model pay trading costs, it ends up favoring value and quality, which is what AVUV and DFSV already hold.
What AVUV and DFSV already do
Both are actively managed, rules-based portfolios that tilt a market-cap-weighted universe toward smaller, cheaper and more profitable companies, and both say in their prospectuses that they weigh trading costs.
| AVUV (Avantis) | DFSV (Dimensional) | |
|---|---|---|
| Net expense ratio | 0.25% | 0.27% since Nov 1, 2026 |
| Assets | $32.5 billion (Nov 13, 2026) | $26.5 billion, whole fund (Sep 30, 2026) |
| Holdings | 790 (Oct 31) | 1,044 (Sep 30) |
| Weighted average market cap | $3.9 billion (Oct 31) | $5.7 billion; Russell 2000 Value $3.3 billion (Sep 30) |
| Turnover, latest fiscal year | 6% | 7% (DFSVX); 9% for the ETF before the merger |
| Value measure | “Adjusted book/price ratio”; book value excludes goodwill | Price to book; may also use price to cash flow or earnings |
| Profitability measure | “Adjusted cash from operations to book value” | Operating profits relative to book value or assets |
| Shorter-term inputs | “Past performance of the security relative to other securities”; the word momentum does not appear | Price momentum, short-run reversals, asset growth |
| Trading and taxes | May consider “expected implementation costs and tax consequences” | “Flexible trading approach” that manages turnover and costs |
Sources: Avantis fund page, January 2026 summary prospectus and white paper; Dimensional fact sheet (September 30, 2026), February 2026 summary prospectuses and the September 25, 2026 fee supplement.1,2,3,5,7,8,9
One change matters for anyone comparing the two. On September 25, 2026, the stand-alone DFSV ETF became an ETF share class of Dimensional’s U.S. Small Cap Value Portfolio, the mutual fund behind ticker DFSVX, which has run since 1993.6 The DFSV ticker now reports that fund’s assets and record. Its fee went from 0.30% for the old ETF to a 0.31% cap for the new share class, then to 0.27% on November 1, 2026.7 The regressions below use DFSVX’s returns for that reason.
Dimensional is still changing its process. On November 4, 2026, it said it “is starting to incorporate profit growth systematically” in every strategy that already targets the profitability premium, based on a 2026 paper by its researchers, and gave no effective date.10,11 A new fund that adds profit growth would be matching Dimensional, and DFSV holders get the change without trading.
The ETF wrapper is also taken. Moussawi, Shen and Velthuis estimate that in-kind redemptions, which let an ETF hand low-basis shares to authorized participants instead of selling them, were worth about 1.05% a year after tax relative to active mutual funds in recent years.12 That is an average for taxable investors measured against mutual funds; inside an IRA or 401(k) it does not apply, and each fund’s actual capital-gains distributions are the real test. AVUV’s prospectus, for example, warns that it may create and redeem shares for cash, which can make it less tax-efficient than an ETF that works entirely in kind.2 DFSV now shares one portfolio, and so its capital-gains distributions, with the mutual fund’s other share classes. A new ETF could use the same structure, and whether it would beat either fund on taxes would show up in its distributions.
Five ways a fund could try to beat them
That leaves five places a challenger could look: a lower fee, deeper or smaller stocks, better signals, momentum, and better trading. For most of them a live fund has made a related bet, so we measured each against AVUV on six factors (the Fama-French five plus momentum) over October 2019 to September 2026, the 84 months all of them overlap.13 Each of these funds differs from AVUV in several ways at once, including its universe, weighting, concentration and fee, so the comparison shows how each fund did; it cannot isolate what any single change is worth. The full table is in the next section.
1. Charge less
The simplest way to beat a fund is to run the same strategy for less. AVUV and DFSV are two basis points apart, so there is little room between them, and a new sponsor carries fixed costs that put a floor under its fee until it gathers assets; our ETF launch-cost guide works through that arithmetic. Cheap small-value funds already exist, but they track plain indexes. Vanguard’s VBR (now named for its Morningstar index) charges 0.05%, loads less on size, value and profitability than AVUV, and returned 10.98% a year over the window against AVUV’s 14.81%. A lower fee only wins if the portfolio stays the same.
2. Go deeper and smaller
The case for a smaller fund rests on capacity. AVUV manages $32.5 billion and Dimensional’s small-value fund $26.5 billion, before counting Dimensional’s targeted-value fund, which buys from a similar universe.1,8,17 A fund with a few hundred million could own the smallest, cheapest companies in meaningful weights, where a giant fund’s position would be either too small to matter or too big to trade.
Bridgeway’s Omni Small-Cap Value strategy is close to that design. Its ETF, BSVO, weights by market cap a portfolio of value stocks “priced cheaply relative to some financial measures of worth,” charges 0.45%, and turned over 19% of its portfolio in its latest fiscal year.15 BSVO took over the record of a tax-managed sibling fund in 2023. Bridgeway’s other sibling, the Omni Small-Cap Value mutual fund (BOSVX), holds 601 stocks and $1.2 billion, charges 0.47%, and has a separate return series back to 2011, so it stands in for the strategy here.16 Against AVUV, BOSVX loaded more on size (+0.08, standard error 0.03) and less on profitability (−0.09, standard error 0.03), with the same value loading. It returned 14.07% a year to AVUV’s 14.81%. Its alpha gap against AVUV over those seven years was +0.57% a year (t-statistic 0.42), and over its own history since September 2011 its six-factor alpha was −0.20% a year (t-statistic −0.21). BOSVX ended up with more size exposure, less profitability and no measurable edge beyond those loadings.
The research is mixed on whether that helps. McLean and Pontiff found that published return predictors earned 58% less after publication, and that the returns that survived were concentrated in illiquid, hard-to-arbitrage stocks.19 That is the case for a small fund that can trade such stocks. Against it, in Hou, Xue and Zhang’s replication of 452 anomalies, 65% failed a 5% significance test once microcaps were kept from dominating, including 96% of the anomalies based on trading frictions.18 And Novy-Marx and Velikov found that few strategies with high turnover keep their premiums after trading costs.32 Owning the smallest stocks gives a fund the chance at those returns and the bill for trading them.
3. Use better signals
A different value measure. Pacer’s CALF holds the 200 small caps with the highest free cash flow relative to enterprise value, weighted by free cash flow, and charges 0.59%.20 In March 2025, S&P announced that the index would widen from 100 to 200 stocks and draw them from a broader small-cap universe, so the regressions below mix the old and new designs.21 CALF’s value loading was 0.12 against AVUV’s 0.55. Cash-flow yield and book-to-market pick different stocks, and CALF also loaded negatively on momentum (−0.14). Its yearly return strayed from AVUV’s by 8.8 percentage points, one standard deviation, and it returned 13.14% a year. CALF is a different strategy with its own case, and these numbers give no sign that it improves on AVUV.
Cash-based profitability. Ball, Gerakos, Linnainmaa and Nikolaev found that operating profitability measured from cash flows beats accrual-based measures and absorbs the accruals anomaly.22 Detzel, Novy-Marx and Velikov found that once trading costs are counted, a Fama-French five-factor model using cash profitability beat the alternatives they tested.23 Avantis already defines profitability with cash from operations.2 Our cash-based profitability guide covers the details.
Intangible assets. Book value leaves out brands, software and research a company built itself. Arnott, Harvey, Kalesnik and Linnainmaa found that adding capitalized intangibles to book value beat traditional value “by a wide margin” and that changes in the valuation spread explain the whole 2007–2020 drawdown of the value factor.24 Eisfeldt, Kim and Papanikolaou report that an intangible-adjusted value factor minus the standard one earned 2.11% a year in their sample, and the two are 76% correlated.25 Dimensional’s researchers, using a similar method, reached the opposite conclusion: they found “no tangible performance benefits from adjusting for intangibles” and attributed the adjusted measure’s better record over the decade to 2018 mainly to different sector exposures.26 Avantis’s published change to book value comes at the same comparison from the other side: it removes goodwill, the intangible a company books when it buys another, saying its profitability measure already counts the acquired firm’s cash flows.3 So the evidence on intangibles is contested. The favorable estimates come from paper portfolios before costs, and they are the kind of public result McLean and Pontiff found shrinking after publication.
One composite score or separate sleeves. Fitzgibbons, Friedman, Pomorski and Serban argue that long-only investors should combine value, momentum and other signals into one score per stock rather than run them as separate portfolios.27 Ghayur, Heaney and Platt found the opposite at low and moderate tracking error: in Russell 1000 value and momentum tests from 1979 to 2016, separate sleeves had an information ratio of 0.89 against 0.59 for one score, and the composite won only at high tracking error.28 The evidence is split, so neither approach is an obvious upgrade.
4. Add momentum
Value and momentum tend to offset each other. Asness, Moskowitz and Pedersen measured a correlation of −0.65 between the two U.S. stock factors.29 Momentum also crashes after market bottoms: from March to May 2009, the past-loser decile rose 163% while past winners gained 8%, a disaster for a strategy short the losers.30
Invesco’s XSVM makes momentum part of the selection. It takes the 240 S&P SmallCap 600 stocks with the best value scores, keeps the 120 with the strongest momentum, weights them by value score, and turned over 84% of its portfolio in its latest year. It has tracked this index since June 2019.31 Its momentum loading was +0.11 with a standard error of 0.11, too noisy to tell from zero, while its yearly return strayed from AVUV’s by 10.3 points. The six factors explained only 7% of that difference; the rest was stock-specific. XSVM returned 13.37% a year.
Dimensional uses momentum differently: its prospectus lets it increase, reduce or skip a holding “based on shorter-term considerations, such as a company’s price momentum, short-run reversals, and investment characteristics.”5 Used that way, momentum keeps a fund from buying a cheap stock while it is still falling or selling a winner the day it stops looking cheap. It does not give the fund momentum exposure: DFSVX’s momentum loading was −0.01 over the window and −0.07 since 1993. Any benefit would show up as lower trading costs, which outsiders cannot measure.
5. Trade better
This is where professional managers spend their effort, and the research says it pays. Novy-Marx and Velikov found that a buy/hold spread, which buys a stock above one rank and sells only below a lower one, was the most effective simple way to cut trading costs, and that strategies with under 50% monthly turnover mostly kept their premiums after costs.32 In a later comparison, trading only within bands of the target beat rebalancing less often, with similar savings and better exposure.33 Frazzini, Israel and Moskowitz, using $1.7 trillion of AQR’s own executions, put real trading costs an order of magnitude below earlier academic estimates.34
AVUV turned over 6% of its portfolio and DFSVX 7% in their latest fiscal years, and both describe cost-aware trading.2,9 A challenger could do it better, but nobody outside the firms can measure by how much, because the saving would be mixed into net returns with everything else.
The scoreboard
The first table shows each fund’s return and its loadings on the size, value, profitability and momentum factors over the common window. A loading of 0.55 on value means the fund moved about 0.55% for each 1% the value factor returned, after accounting for the other factors.
| Fund | Idea | ETF fee | Return/yr | Worst drop | Market | Size | Value | Profit. | Mom. | Alpha/yr (t) |
|---|---|---|---|---|---|---|---|---|---|---|
| AVUV | Baseline | 0.25% | 14.81% | −42.4% | 1.06 | 0.89 | 0.55 | 0.22 | 0.01 | +0.51 (+0.39) |
| DFSVDFSVX returns | Baseline | 0.27% | 13.70% | −39.0% | 0.99 | 0.80 | 0.50 | 0.21 | −0.01 | +0.04 (+0.04) |
| DFATDFFVX returns | Wider, adds mid caps | 0.26% | 13.41% | −39.2% | 0.99 | 0.78 | 0.50 | 0.18 | −0.02 | −0.24 (−0.25) |
| BSVOBOSVX returns | Deeper value, smaller stocks | 0.45% | 14.07% | −41.8% | 0.98 | 0.97 | 0.56 | 0.13 | 0.03 | +1.09 (+0.64) |
| CALF | Free-cash-flow yield | 0.59% | 13.14% | −35.1% | 0.99 | 0.89 | 0.12 | 0.23 | −0.14 | +0.54 (+0.16) |
| XSVM | Value plus momentum | 0.36% | 13.37% | −41.0% | 1.06 | 1.05 | 0.41 | 0.22 | 0.11 | −0.27 (−0.10) |
| VBR | Plain index | 0.05% | 10.98% | −35.0% | 0.98 | 0.60 | 0.39 | 0.12 | −0.01 | −2.87 (−3.05) |
| IWN | Plain index | 0.24% | 10.42% | −35.8% | 0.98 | 0.83 | 0.40 | −0.01 | −0.00 | −2.31 (−2.57) |
Summitward calculation, October 2019 to September 2026 (84 months). Returns are compound annual, net of fees. Loadings come from regressing each fund’s monthly return above the T-bill rate on the Fama-French five factors plus momentum (investment loadings are in the research output); t-statistics use Newey-West standard errors. Returns and loadings for DFSV and DFAT come from DFSVX and DFFVX, older share classes of the same funds, and for BSVO from BOSVX, its sibling mutual fund, each net of its own fees; the fee column shows what each ETF charges today. Those fees were higher for most of the window: DFSVX’s fell from 0.51% (February 2020) to 0.31% (February 2026), DFFVX’s from 0.36% to 0.29%, and BOSVX’s from 0.62% (October 2019) to 0.47%, and the two Dimensional classes dropped to their ETF classes’ fees on November 1, 2026. AVUV charges 0.25%.4,7,9,16,17 Data: Ken French library (202609 CRSP database), Yahoo Finance adjusted closes.13
The second table puts each fund directly against AVUV: how far its yearly return typically strayed from AVUV’s (tracking error), the gap in compound return, the alpha left after the six factors, and the edge its extra or missing loadings would imply if each style factor earned its 1963–2026 average, after subtracting the fee difference.
| Fund minus AVUV | Tracking error | Return gap/yr | Alpha gap/yr (t) | Implied edge, full premiums | Stylized: 42% of premiums |
|---|---|---|---|---|---|
| DFSVDFSVX returns | 3.5% | −1.12 | −0.47 (−0.51) | −0.46 [−1.0, +0.1] | −0.20 |
| DFATDFFVX returns | 3.6% | −1.40 | −0.76 (−0.85) | −0.72 [−1.2, −0.2] | −0.31 |
| BSVOBOSVX returns | 3.8% | −0.75 | +0.57 (+0.42) | +0.12 [−0.5, +0.7] | −0.07 |
| CALF | 8.8% | −1.68 | +0.02 (+0.01) | −2.23 [−4.7, +0.3] | −1.13 |
| XSVM | 10.3% | −1.44 | −0.79 (−0.29) | +1.28 [−1.9, +4.4] | +0.47 |
| VBR | 6.6% | −3.83 | −3.38 (−2.57) | −1.35 [−2.5, −0.2] | −0.45 |
| IWN | 5.9% | −4.39 | −2.83 (−1.93) | −1.41 [−2.5, −0.3] | −0.58 |
Summitward calculation, October 2019 to September 2026. Tracking error is the annualized standard deviation of the monthly return difference. The alpha gap comes from regressing that difference on the six factors. Implied edge is the loading gap on size, value, profitability, investment and momentum times each factor’s July 1963 to September 2026 average, minus the fee difference, with a 95% interval from both the premium and loading estimates. The market loading is left out, because extra market exposure can be added by holding more of any fund. The last column is a stylized sensitivity: it scales every style premium to 42%, applying McLean and Pontiff’s average post-publication decline across 97 published predictors to each factor alike, which their study did not estimate factor by factor, and the 1963–2026 averages already include years after each factor was published, so the haircut partly counts the decline twice. If the style premiums were zero from here on, each implied edge would reduce to AVUV’s fee minus the fund’s. The intervals assume the six-factor model and the 1963–2026 averages; the columns show what the loadings imply under those premiums and forecast nothing.13,19
Every fund trailed AVUV, and for the five small-value alternatives none of the gaps can be told from luck: their alpha gaps run from −0.79% to +0.57% a year, with t-statistics between −0.85 and +0.42. For the Dimensional, Bridgeway and Pacer funds, half or more of AVUV’s lead came from one loading. AVUV’s market loading of 1.06 compared with their 0.98 to 0.99, and the market returned 13.6% a year above T-bills over the window. After the fact, that difference alone accounts for 0.9 to 1.1 points a year of their arithmetic return gaps, which ran from 0.8 to 1.9 points. XSVM’s market loading matched AVUV’s, so its gap came from elsewhere. The size factor lost 1.9% a year over the same months, so BOSVX, which loaded more on size than AVUV, gave some back there. Fees account for part of the rest: for most of the window the Dimensional and Bridgeway funds charged more than AVUV’s 0.25%, and more than their ETFs charge today.
AVUV is the yardstick here because it is the fund the question is about, and it had a good window. Measured against the six factors instead, with no fund as the baseline, AVUV’s own alpha was +0.51% a year (t-statistic 0.39), so each fund’s alpha gap against AVUV is its own alpha minus that half point. Changing the yardstick, to the factors or to DFSV, moves the zero line and keeps the order: BOSVX (+1.09%, t 0.64) and CALF (+0.54%) sit above AVUV, DFSVX (+0.04%) near zero, and DFFVX and XSVM slightly below it. All six alphas stay within the range luck would produce. The chart shows the three ways of measuring side by side.
How Far Apart, and How Sure?
Dots are estimates and bars are 95% intervals, in percentage points a year, October 2019 to September 2026.
Six-factor alpha of each fund's monthly return minus AVUV's. Left of zero means behind AVUV after accounting for factor exposure.
- DFSV (DFSVX returns)−0.47 [−2.3, +1.4]
- DFAT (DFFVX returns)−0.76 [−2.5, +1.0]
- BSVO (BOSVX returns)+0.57 [−2.1, +3.3]
- CALF+0.02 [−5.6, +5.7]
- XSVM−0.79 [−6.2, +4.6]
- VBR−3.38 [−6.0, −0.8]
- IWN−2.83 [−5.7, +0.05]
Summitward calculation, October 2019 to September 2026 (84 months), from Yahoo Finance adjusted closes and Ken French factors. Alpha intervals use Newey-West standard errors (six lags); the implied-edge interval combines uncertainty in the 1963 to 2026 premiums and in the loadings. All intervals assume the six-factor model; they describe a past window and forecast nothing. The DFSV and DFAT rows use older share classes of the same funds; the BSVO row uses BOSVX, a sibling Bridgeway fund.
On loadings alone, two challengers imply edges over AVUV and none is clearly ahead. XSVM’s loadings imply +1.28% a year at full premiums, but the 95% interval runs from −1.9 to +4.4 points because its investment and momentum loadings are measured so loosely. BOSVX’s gains and losses across factors net to about zero (+0.12, interval −0.5 to +0.7) after its higher fee. CALF’s lower value and momentum loadings imply about 2.2 points a year less than AVUV (−4.7 to +0.3), and DFSVX’s slightly lower loadings about 0.5 points less (−1.0 to +0.1). DFFVX’s lower loadings on size, value, profitability and momentum imply about 0.7 points a year less (−1.2 to −0.2), the one small-value challenger whose interval stays below zero.
Among the small-value funds, the largest difference separates AVUV and DFSV from the two plain index funds. VBR and IWN trailed AVUV by 3.8 and 4.4 points a year and DFSVX by 2.7 and 3.3. Their loadings imply about 1.4 points a year less than AVUV’s at full premiums, with 95% intervals that stay below zero (−2.5 to −0.2 for VBR, −2.5 to −0.3 for IWN). Their own six-factor alphas were also negative, −2.87% and −2.31% a year, but index funds tend to show negative alphas in Fama-French models for reasons that have nothing to do with management, so the loadings are the sturdier evidence.35 The gap fits the case for targeting value and profitability deliberately, but one seven-year window cannot show how much of it came from the profitability screen, from trading, or from other differences between the portfolios.
A note on the data: Dimensional’s targeted-value fund (DFFVX, now also DFAT) adds mid caps, which shows up as a lower size loading. The AVUV vs. BSVO vs. DFSV comparison covers which of the three fits which investor.
Can machine learning do better?
The strongest academic case for machine learning in stock selection is Gu, Kelly and Xiu. Across about 30,000 U.S. stocks from 1957 to 2016 and 94 stock characteristics, a three-layer neural network predicted monthly returns best, with an out-of-sample R² of 0.40%. Every method agreed on which inputs mattered most: recent price trends, including the one-month reversal, then liquidity, then volatility. Their value-weighted neural-network portfolio turned over about 123% a month.36
Those are the signals a low-turnover, long-only fund avoids. Avramov, Cheng and Metzker reran these models with the restrictions a real fund faces. Leaving out microcaps cut the Gu-Kelly-Xiu-style alpha by 66%, to an insignificant 0.31% a month, and leaving out financially distressed firms cut it by 78%. The strategies turned over at least 87% of the portfolio each month.37 Chen, Pelger and Zhu’s deep-learning model posted an annual Sharpe ratio of 2.6 before costs on the full universe, falling to 0.9 when restricted to the 550 largest stocks.39
The newest result speaks directly to this question. Jensen, Kelly, Malamud and Pedersen trained machine-learning portfolios with trading costs built into the objective, for an investor with $10 billion, over 1981 to 2020. A model that predicted returns well but ignored costs had a Sharpe ratio of 2.06 before costs and deeply negative returns after them, with enormous turnover. Their cost-aware version earned a net Sharpe ratio of 1.33, and the characteristics it relied on most were value and quality, where the naive models leaned on short-term reversal.38 These are long-short portfolios, so a long-only ETF could not match that Sharpe ratio, but the shift in signals applies to any portfolio: a model that pays to trade moves toward the slow signals AVUV and DFSV are built on.
The live record is thin. AIEQ, launched in October 2017 and run on EquBot’s system on IBM Watson, is the longest-running ETF sold on machine-learning stock selection; it was actively managed with EquBot as sub-adviser until January 2024 and has tracked an EquBot index since. It charges 0.75% and turned over 804% of its portfolio in its latest fiscal year.40
| AIEQ period | Months | AIEQ return/yr | VTI return/yr | Six-factor alpha/yr (t) |
|---|---|---|---|---|
| Active, EquBot sub-adviser (2017-11 to 2024-01) | 75 | 7.65% | 11.86% | −1.03 (−0.35) |
| EquBot index (2024-02 to 2026-09) | 32 | 14.62% | 19.67% | −4.46 (−0.99) |
| Full history (2017-11 to 2026-09) | 107 | 9.69% | 14.14% | −2.42 (−1.04) |
Summitward calculation from Yahoo Finance adjusted closes; factors from the Ken French library. AIEQ is an all-cap fund, so VTI is the natural comparison. Amplify reports a 9.37% annualized NAV return since inception through October 31, 2026.13,41
AIEQ trailed the total market in both eras, by about 4.4 points a year over its whole life. Even so, the arithmetic gap to VTI, about 3.2 points a year, has a t-statistic of only −1.0, because the fund’s return strayed from VTI’s by 9.5 points a year; its six-factor alpha of −2.42% a year is no more conclusive (t −1.04). AIEQ’s record tests one marketing claim, that an AI-run fund beats the market; it says little about whether cost-aware machine learning would improve a small-value fund. Studies of AI funds as a group find no edge over the market either. Chen and Ren found that 15 AI-powered mutual funds beat human-managed peers over 26 months but did not beat the market, with an alpha that was not statistically significant. The European Securities and Markets Authority found that EU funds using AI had not delivered significantly higher or lower performance than comparable funds.42,43 I found no U.S. small-cap fund run on machine learning with three years of history, and no public statement that Avantis or Dimensional uses machine learning to build portfolios.
Where machine learning can plausibly help a small-value fund is in the unglamorous parts: forecasting the cost of each trade, combining slow signals in ways a linear score misses, and catching bad data. How much any of them would add over what Avantis and Dimensional already do has not been measured, and like better trading, the gain would be hard to see from outside.
What about large language models and AI agents?
The best-known paper here has been withdrawn. In 2024, Kim, Muhn and Nikolaev reported that GPT-4, reading anonymized financial statements, predicted the direction of next year’s earnings with 60% accuracy against 53% for analysts. On February 20, 2025, the authors withdrew it from circulation, pending a review, after a co-author found inconsistencies in the data while trying to replicate the results.44
Newer work is more modest. Dou and Fu fine-tuned a three-billion parameter open model on financial statements to predict earnings changes.45 In a November 2026 comparison on 2023 and 2024 data, deHaan and Huang measured that model at an AUC of 67.88%, where 50% is a coin flip, against 69.53% for a tabular foundation model and 62.75% to 63.78% for tree-based models.46 That is better than chance at predicting earnings direction. No study I found shows it surviving in a diversified, long-only portfolio after costs.
Backtesting a language model has a problem of its own: the model may have read the future. Lopez-Lira, Tang and Zhu found that GPT-4o recalled S&P 500 index levels within 0.61% on average for dates before its training cutoff and missed by 16.87% after it, and removing company names did not stop the recall.47 The size of the problem varies. Didisheim, Fraschini and Somoza found the bias negligible for smaller models and finer-grained data, and Glasserman and Lin found that company names distracted the model more than they leaked the future: anonymized headlines did better than named ones inside the training window.48,49 A clean test needs models trained only on information available at each test date, which shortens the usable history.
So does the data. Smaller public companies began tagging their financial statements in machine-readable XBRL with fiscal periods ending after June 15, 2011.50 For earlier years, a builder has two choices: license a point-in-time database, or have a model pull the numbers out of the text filings on EDGAR, where every domestic public company has filed electronically since May 1996.51 Any extraction errors then become part of the backtest. And agents still make errors on financial research: in the 2026 AI Index, the best model answered 63.33% of a finance-analyst benchmark’s questions correctly.52
The realistic job for an agent is checking work: reading footnotes to flag one-time charges that distort a profitability ratio, catching data errors before they reach the portfolio, and tracking corporate actions, with rule-based checks deciding what goes into the model. That could make a fund cheaper to run, and catching a wrong number before it moves a profitability score can change what the fund buys, so it could affect returns as well. Neither effect has been measured. A clean test would run the same portfolio rules on the raw data and on the agent-checked data and compare the results. Our guide to agentic trading accounts covers the retail version of the same question.
How long until you would know
Suppose a new fund really does beat AVUV by some amount a year, and its yearly return differs from AVUV’s with standard deviation , its tracking error. After years, the average difference has a standard error of . For the edge to reach a t-statistic of about 2 on average, you need:
DFSVX, the closest substitute for AVUV, had 3.5% tracking error against it. A true 1%-a-year edge at that tracking error would take about 48 years to reach that point, and a 0.5% edge about 190 years. Even then a test would confirm the edge only about half the time; an 80% chance of confirming a 1% edge takes about 98 years. At CALF’s 8.8% the 1% edge would take about 300 years to reach a t-statistic of 2, and at XSVM’s 10.3% about 410. After seven years, AVUV’s entire life so far, a fund that truly beat it by 1% a year with 3.5% tracking error would be ahead about 77% of the time, and a standard test would confirm the edge about 11% of the time. Netting out the six factors lowers the tracking error, but only to 2.9% for DFSVX, which still means about 33 years.
A backtest cannot shortcut the wait. A designer who tries many variations and keeps the best one inflates its historical result, and Bailey, Borwein, López de Prado and Zhu show how often the in-sample winner then underperforms out of sample.54 The factor research itself is contested: Hou, Xue and Zhang found most anomalies failing replication, while Jensen, Kelly and Pedersen, with a different method, found 82% of factors replicating.18,53 Value and profitability are among the best-supported factors in either camp, which is a reason to own them, but no backtest can establish that one implementation beats another by half a point.
That wait applies to proving an edge from a fund’s live returns. Other evidence comes faster, because it can be tested across thousands of stocks or trades at once: a manager’s trading-cost records, forecasts of individual companies’ earnings, or the same rules run on two versions of the data. None of that proves a fund’s future return, but it can show whether a change works as designed.
The recipe is public
If I had to design a small-cap value fund from the research above, it would look like this:
- Start with U.S. small caps, market-cap weighted, with only a light reach into microcaps.
- Tilt toward stocks cheap on book value, cross-checked with cash-flow and earnings yields, and test an intangibles adjustment.
- Tilt toward profitable companies, measured from cash flows. Novy-Marx found that controlling for gross profitability improves value strategies, and Ball and coauthors found cash-based measures beat accrual-based ones.22,55
- Avoid small companies that grow assets quickly while earning little, the five-factor model’s biggest failure.56
- Use momentum to time trades: hold off buying falling stocks and selling rising ones.
- Trade only when a stock moves well outside a band around its target, keep turnover in single digits, and use an ETF wrapper.
Most of that list appears, in some form, in the AVUV or DFSV prospectus. The intangibles test does not appear in either firm’s published materials, and Dimensional has tested it and argued against adopting it. What a newcomer cannot copy is the rest: the fee that $30 billion of assets supports, decades of trading data, and the details of how each firm turns those rules into orders. Nobody outside the firms can measure what those details are worth.
What I recommend
If you want a U.S. small-cap value tilt, own AVUV or DFSV, or both. None of the alternatives here has shown a better result after costs and risk, which is short of proof that the two cannot be improved. Whether you want the tilt at all depends on the rest of your portfolio and on whether you can hold through years of trailing the market. The two funds are run by different firms with different rules, which spreads implementation risk and makes them candidates to pair for tax-loss harvesting. The IRS has never ruled on when two funds are substantially identical, so that pairing is a judgment call; our tax-loss harvesting guide covers it. The AVUV vs. BSVO vs. DFSV comparison covers how the two differ.
If you want a bigger tilt, hold more of the fund you have. A larger small-value share of your portfolio raises your size, value and profitability exposure together, along with market exposure (AVUV’s market loading was 1.06), at a 0.25% fee and single-digit turnover. The deeper and more concentrated funds in the table changed the mix more than the dose: BOSVX added size and gave up profitability, while XSVM’s loadings were measured too loosely to pin down. Choose one of them for its specific mix, and expect more stock-specific risk. Our guide on whether higher loadings mean higher expected returns shows how uncertain that extra premium is.
Treat any fund sold as a better or AI-powered AVUV as unproven until it has a long live record. Ask for its factor loadings, turnover, fee, assets and live returns, and discount backtests. A fund whose main evidence is a backtest is asking you to pay for a result that, by the arithmetic above, nobody can verify.
This advice would change if a large sponsor launched the same strategy at a clearly lower fee with matching loadings; if AVUV’s size began to show in its characteristics, such as a weighted market cap rising relative to its benchmark or a shrinking value spread; or if a challenger built a decade or more of positive alpha against AVUV at low tracking error.
Key Takeaways
- AVUV and DFSV already screen on profitability, manage trading costs, and use the ETF wrapper’s tax advantages. Dimensional uses momentum to time trades and said on November 4, 2026 that it is adding profit growth.
- From October 2019 to September 2026, four funds built around different ideas (wider targeted value, deeper value, cash-flow yield, momentum) trailed AVUV by 0.7 to 1.7 points a year, for three of them about half or more of it from AVUV’s higher market loading. None of the six-factor alpha gaps had a t-statistic above 1 in absolute value, and measured against the six factors alone, BOSVX’s alpha was higher than AVUV’s.
- Among the small-value funds, the largest gap separates the plain index funds from AVUV and DFSV: over the same window VBR and IWN trailed AVUV by 3.8 and 4.4 points a year, and their loadings alone imply about 1.4 points a year less, with intervals below zero. One window cannot say which design choice produced the gap.
- Machine learning’s published gains rely on high-turnover signals in small and distressed stocks; with trading costs built in, the best model leaned on value and quality. AIEQ, an all-cap AI-run fund, trailed VTI by about 4.4 points a year from November 2017 to September 2026.
- At DFSVX’s 3.5% tracking error against AVUV, a true 1%-a-year edge would take about 48 years of live returns just to reach a t-statistic of 2 on average, and about 98 years for an 80% chance of confirming it.
Frequently Asked Questions
Is there a better small-cap value ETF than AVUV?
Not one you can identify from the data so far. Over October 2019 to September 2026, DFSVX, DFFVX, BOSVX, CALF and XSVM all returned less than AVUV, but none of the differences is statistically meaningful. For DFSVX, DFFVX, BOSVX and CALF, half or more of AVUV’s lead came from its higher market loading in a strong market. Measured against the six factors instead of AVUV, BOSVX’s alpha was higher than AVUV’s, by a margin too small to tell from luck. DFSV is the closest alternative.
Is DFSV the same fund as DFSVX now?
Since September 25, 2026, DFSV has been an ETF share class of the same Dimensional fund as DFSVX, so the two hold the same portfolio. The ETF class charges 0.27% since November 1, 2026.
Can machine learning pick small-cap value stocks better than Avantis or Dimensional?
There is no live evidence that it can. Academic machine-learning portfolios earn most of their returns from fast signals in microcaps and distressed stocks; the best-known model’s alpha fell by two-thirds when microcaps were excluded. A 2026 study that trained the model with trading costs found it relied most on value and quality, the signals these funds already use.
Do Avantis or Dimensional use AI or machine learning?
Neither firm’s prospectus or public research says it uses machine learning to build portfolios, as of November 2026. Both describe rules-based selection on value and profitability with cost-aware trading.
How many years does it take to know whether a fund beats AVUV?
A true edge takes roughly (1.96 × tracking error ÷ edge) squared years to reach a t-statistic of about 2 on average, and a test at that point confirms it only half the time. With 3.5% tracking error against AVUV, a 1%-a-year edge needs about 48 years and a 0.5% edge about 190 years. More concentrated funds, with 8% to 10% tracking error, need centuries to show a 1% edge.
Did the AI-powered ETF AIEQ beat the market?
No. From November 2017 to September 2026, AIEQ returned 9.69% a year against 14.14% for VTI, by our calculation from adjusted closes, and trailed in both its actively managed and index-tracking periods.
Related Guides
- AVUV vs BSVO vs DFSV Which small-cap value ETF fits which investor.
- Alpha Architect ETFs The case for concentrated factor funds, with regressions.
- Do Higher Factor Loadings Mean Higher Expected Returns? How uncertain a loading’s implied premium is.
- Is Quant Investing Just Factor and Risk Models? The risk-model side of quant funds.
- RMW Explained The profitability factor and the Avantis-Dimensional difference.
- What It Actually Costs to Launch an ETF Why a small sponsor cannot undercut a quarter-point fee.
- Is the Value Premium Back? What drove small value’s 2026 run.
- Robinhood’s Agentic Trading The evidence on letting an AI agent trade for you.
Sources
- Avantis Investors, Avantis U.S. Small Cap Value ETF fund page: expense ratio 0.25%, net assets $32.47 billion (NAV dated November 13, 2026), 790 holdings and weighted average market cap $3.91 billion at October 31, 2026. avantisinvestors.com
- American Century ETF Trust, Avantis U.S. Small Cap Value ETF summary prospectus (497K), January 1, 2026: strategy text and 6% turnover for the fiscal year ended August 31, 2025. sec.gov
- Avantis Investors, “Our Scientific Approach to Investing,” white paper, 2026: the modified book-to-market measure excludes goodwill. americancentury.com
- Fees at the start of the window: DFA U.S. Small Cap Value Portfolio (DFSVX) and U.S. Targeted Value Portfolio (DFFVX) summary prospectuses, February 28, 2020, total annual operating expenses 0.51% and 0.36% sec.gov, sec.gov; Bridgeway Omni Small-Cap Value Fund (BOSVX) prospectus, October 31, 2019, 0.62% after fee waiver. sec.gov
- Dimensional ETF Trust, Dimensional U.S. Small Cap Value ETF summary prospectus (497K), February 28, 2026: strategy text, 9% turnover, 0.30% total expenses. sec.gov
- Dimensional, supplement (497) of June 18, 2026 on reorganizing DFSV into an ETF share class of the U.S. Small Cap Value Portfolio, expected on or around September 25, 2026. sec.gov
- DFA Investment Dimensions Group, supplement (497) dated September 25, 2026: ETF Class expense cap of 0.31%, falling to 0.27% on November 1, 2026, when the Institutional Class limits fall to 0.27% (U.S. Small Cap Value) and 0.26% (U.S. Targeted Value). sec.gov
- Dimensional, U.S. Small Cap Value ETF fact sheet, September 30, 2026: $26.5 billion, 1,044 holdings, weighted average market cap $5,669 million versus $3,325 million for the Russell 2000 Value Index. dimensional.com
- DFA Investment Dimensions Group, U.S. Small Cap Value Portfolio summary prospectus (497K), February 28, 2026: 7% turnover, 0.31% total expenses. sec.gov
- Namiko Saito and Stanley Black, “Q&A on Profit Growth and Expected Returns,” Dimensional, November 4, 2026. dimensional.com
- Ahn et al., “Profit Growth and Expected Stock Returns,” Dimensional working paper, 2026. ssrn.com
- Rabih Moussawi, Ke Shen and Raisa Velthuis, “The Role of Taxes in the Rise of ETFs,” Review of Financial Studies 38(10), 2025, 2988–3039. doi.org
- Summitward calculations: monthly returns of AVUV, DFSVX, DFFVX, BOSVX, CALF, XSVM, VBR, IWN and AIEQ, October 2019 to September 2026 (84 months), in excess of the T-bill rate and as differences from AVUV, regressed on the Fama-French five factors plus momentum with Newey-West standard errors; loading gaps times 1963–2026 factor averages; each fund’s own history by regime. Data: Ken French data library (202609 CRSP database), Yahoo Finance adjusted closes. Script and output: summitward-research
- Avantis U.S. Small Cap Value ETF fact sheet and Dimensional U.S. Small Cap Value Portfolio (Institutional Class) fact sheet, September 30, 2026: five-year annualized NAV returns of 11.17% and 10.97%. avantisinvestors.com, dimensional.com
- EA Series Trust, EA Bridgeway Omni Small-Cap Value ETF summary prospectus (497K), October 31, 2026: 0.45% expenses, 19% turnover for the year ended June 30, 2026, reorganized from the Bridgeway Omni Tax-Managed Small-Cap Value Fund on March 10, 2023. sec.gov
- Bridgeway Funds, Omni Small-Cap Value Fund page: 0.47% net expense ratio, $1.2 billion and 601 holdings at September 30, 2026, inception August 31, 2011. bridgewayfunds.com
- DFA Investment Dimensions Group, U.S. Targeted Value Portfolio summary prospectus (497K), 2026: companies smaller than the 500th largest U.S. company, 10% turnover, 0.29% net expenses. sec.gov
- Kewei Hou, Chen Xue and Lu Zhang, “Replicating Anomalies,” Review of Financial Studies 33(5), 2020, 2019–2133. doi.org
- R. David McLean and Jeffrey Pontiff, “Does Academic Research Destroy Stock Return Predictability?” Journal of Finance 71(1), 2016, 5–32. doi.org
- Pacer Funds Trust, Pacer US Small Cap Cash Cows ETF summary prospectus (497K), August 31, 2026: top 200 stocks by free cash flow yield, 0.59% expenses, 130% turnover. sec.gov
- S&P Dow Jones Indices, “Pacer US Small Cap Cash Cows Index: consultation results,” March 11, 2025 (constituents raised from 100 to 200). spglobal.com
- Ray Ball, Joseph Gerakos, Juhani Linnainmaa and Valeri Nikolaev, “Accruals, Cash Flows, and Operating Profitability in the Cross Section of Stock Returns,” Journal of Financial Economics 121(1), 2016, 28–45. doi.org
- Andrew Detzel, Robert Novy-Marx and Mihail Velikov, “Model Comparison with Transaction Costs,” Journal of Finance 78(3), 2023, 1743–1775. doi.org
- Robert Arnott, Campbell Harvey, Vitali Kalesnik and Juhani Linnainmaa, “Reports of Value’s Death May Be Greatly Exaggerated,” Financial Analysts Journal 77(1), 2021, 44–67. doi.org
- Andrea Eisfeldt, Edward Kim and Dimitris Papanikolaou, “Intangible Value,” Critical Finance Review 11(2), 2022, 299–332; figures from NBER Working Paper 28056 (November 2021 revision). doi.org, nber.org
- Savina Rizova and Namiko Saito, “Untangling Intangibles,” Dimensional Fund Advisors, July 27, 2021 dimensional.com; Dimensional Fund Advisors, “The Right Thing, Not the Easy Thing,” November 16, 2020. dimensional.com
- Shaun Fitzgibbons, Jacques Friedman, Lukasz Pomorski and Laura Serban, “Long-Only Style Investing: Don’t Just Mix, Integrate,” Journal of Investing 26(4), 2017, 153–164. doi.org
- Khalid Ghayur, Ronan Heaney and Stephen Platt, “Constructing Long-Only Multifactor Strategies: Portfolio Blending vs. Signal Blending,” Financial Analysts Journal 74(3), 2018, 70–85; information ratios as summarized by CFA Institute. doi.org, cfainstitute.org
- Clifford Asness, Tobias Moskowitz and Lasse Heje Pedersen, “Value and Momentum Everywhere,” Journal of Finance 68(3), 2013, 929–985, Table I. doi.org
- Kent Daniel and Tobias Moskowitz, “Momentum Crashes,” Journal of Financial Economics 122(2), 2016, 221–247; figures from NBER Working Paper 20439. doi.org
- Invesco Exchange-Traded Fund Trust, Invesco S&P SmallCap Value with Momentum ETF summary prospectus (497K), August 28, 2026: index method, 0.36% expenses, 84% turnover, prior index used through June 21, 2019. sec.gov
- Robert Novy-Marx and Mihail Velikov, “A Taxonomy of Anomalies and Their Trading Costs,” Review of Financial Studies 29(1), 2016, 104–147. doi.org
- Robert Novy-Marx and Mihail Velikov, “Comparing Cost-Mitigation Techniques,” Financial Analysts Journal 75(1), 2019, 85–102. doi.org
- Andrea Frazzini, Ronen Israel and Tobias Moskowitz, “Trading Costs,” AQR working paper, 2018. aqr.com
- Martijn Cremers, Antti Petajisto and Eric Zitzewitz, “Should Benchmark Indices Have Alpha? Revisiting Performance Evaluation,” Critical Finance Review 2(1), 2013, 1–48. doi.org
- Shihao Gu, Bryan Kelly and Dacheng Xiu, “Empirical Asset Pricing via Machine Learning,” Review of Financial Studies 33(5), 2020, 2223–2273; Table 1 and Table 8 of NBER Working Paper 25398. doi.org
- Doron Avramov, Si Cheng and Lior Metzker, “Machine Learning vs. Economic Restrictions: Evidence from Stock Return Predictability,” Management Science 69(5), 2023, 2587–2619. doi.org
- Theis Ingerslev Jensen, Bryan Kelly, Semyon Malamud and Lasse Heje Pedersen, “Machine Learning and the Implementable Efficient Frontier,” Review of Financial Studies 39(10), 2026, 3035–3078, Table 2. doi.org
- Luyang Chen, Markus Pelger and Jason Zhu, “Deep Learning in Asset Pricing,” Management Science 70(2), 2024, 714–750. doi.org
- Amplify ETF Trust, Amplify AI Powered Equity ETF summary prospectus (497K), January 28, 2026: EquBot index, IBM Watson, 804% turnover. sec.gov; ETF Managers Group summary prospectus, 2023 (actively managed, EquBot sub-adviser). sec.gov
- Amplify ETFs, AIEQ fund page: 0.75% expense ratio, $110.8 million net assets (November 13, 2026), 9.37% annualized NAV return since inception to October 31, 2026. amplifyetfs.com
- Rui Chen and Jinjuan Ren, “Do AI-Powered Mutual Funds Perform Better?” Finance Research Letters 47, 2022, 102616. doi.org
- European Securities and Markets Authority, “Artificial Intelligence in EU Investment Funds: Adoption, Strategies and Portfolio Exposures,” TRV Risk Analysis, February 2025. esma.europa.eu
- Alex Kim, Maximilian Muhn and Valeri Nikolaev, “Financial Statement Analysis with Large Language Models,” arXiv 2407.17866, v2 November 2024; withdrawn in v3, February 20, 2025. arxiv.org
- Yiwei Dou and Qishen Fu, “Supervised Fine-tuning of Large Language Models on Financial Statements to Predict Future Earnings Changes,” NYU working paper. ssrn.com
- Ed deHaan and Allen H. Huang, “Tabular Foundation Models in Financial Predictions,” arXiv 2611.13550, November 2026. arxiv.org
- Alejandro Lopez-Lira, Yuehua Tang and Mingyin Zhu, “The Memorization Problem: Can We Trust LLMs’ Economic Forecasts?” arXiv 2504.14765, v2 December 2025. arxiv.org
- Antoine Didisheim, Martina Fraschini and Luciano Somoza, “AI’s Predictable Memory in Financial Analysis,” Economics Letters 256, 2025, 112602. doi.org
- Paul Glasserman and Caden Lin, “Assessing Look-Ahead Bias in Stock Return Predictions Generated by GPT Sentiment Analysis,” arXiv 2309.17322, 2023. arxiv.org
- U.S. Securities and Exchange Commission, “Interactive Data to Improve Financial Reporting,” Release 33-9002, 2009 (XBRL phase-in by filer size, fiscal periods ending after June 15, 2009, 2010 and 2011). sec.gov
- U.S. Securities and Exchange Commission, “About EDGAR”: phase-in to mandatory electronic filing ended May 6, 1996. sec.gov
- Stanford Institute for Human-Centered AI, AI Index Report 2026, Chapter 2 (Finance Agent v1.1 results from Vals.ai). hai.stanford.edu
- Theis Ingerslev Jensen, Bryan Kelly and Lasse Heje Pedersen, “Is There a Replication Crisis in Finance?” Journal of Finance 78(5), 2023, 2465–2518. doi.org
- David Bailey, Jonathan Borwein, Marcos López de Prado and Qiji Jim Zhu, “The Probability of Backtest Overfitting,” Journal of Computational Finance 20(4), 2017, 39–69. doi.org
- Robert Novy-Marx, “The Other Side of Value: The Gross Profitability Premium,” Journal of Financial Economics 108(1), 2013, 1–28. doi.org
- Eugene Fama and Kenneth French, “A Five-Factor Asset Pricing Model,” Journal of Financial Economics 116(1), 2015, 1–22. doi.org
Method
Fund returns are month-end changes in Yahoo Finance dividend-adjusted closes; factor and T-bill returns are from Ken French’s library, files built from the 202609 CRSP database. The common window, October 2019 to September 2026, starts with AVUV’s first full month. Regressions are ordinary least squares with an intercept on six factors (market, size, value, profitability, investment and momentum), with Newey-West standard errors at six lags; every interval assumes this six-factor model. Maximum drawdown runs from the start of each window. The fee column shows what each ETF charges today (DFSV, DFAT, BSVO), while returns come from DFSVX and DFFVX (older share classes of the same funds) and BOSVX (BSVO’s sibling mutual fund), net of their own fees, which were higher for most of the window. Implied edges multiply each fund’s loading gap against AVUV by the factor’s average from July 1963 to September 2026, leave out the market term, and subtract the fee difference; the interval combines the Newey-West uncertainty of the premiums (12 lags) and of the loading gap. Own-history regressions split AIEQ at its January 2024 move from active management to an index and start XSVM at its June 2019 index change. As a pipeline check, the script reproduces the AVUV loadings published in our factor-loadings guide to three decimal places. As a check on the price data, five-year annualized returns to September 30, 2026 computed this way, 11.15% for AVUV and 11.00% for DFSVX, are within 0.03 points of the NAV returns on the funds’ fact sheets.14 Code, data and printed output are in the summitward-research repository.
Author disclosure
I hold AVUV and other Avantis funds. Summitward has no business relationship with Avantis, Dimensional, Bridgeway, Pacer, Invesco, Amplify, Vanguard, iShares 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.