Why "Factor Investing" Still Works: Applying Fama-French Models in the AI Era

On paper, AI should have arbitraged the Fama-French factors away by now. Hedge funds have spent thirty years feeding the same SMB and HML spreadsheets into ever-faster models, the academic literature has been openly published since 1993, and the cost of running a factor regression has fallen to essentially zero. By the textbook's own logic — once a return-generating pattern is widely known, capital floods in until it disappears — small-value, momentum, and quality should be sitting at zero excess return today. The data tells a more complicated story, and that story is what a long-term investor needs to understand before deciding whether to tilt a portfolio toward any of these factors at all.

What the Fama-French model actually claims

Eugene Fama and Kenneth French's 1992-1993 work isn't a stock-picking system. It's a regression. They observed that single-factor CAPM — which says expected return is determined entirely by market beta — leaves a substantial portion of the cross-section of returns unexplained, and they showed that adding two more variables (Size and Value) explains a large part of that residual. The 2015 five-factor extension added Profitability and Investment.

The crucial subtlety is interpretive. Fama himself argued the factor premia are compensation for risk, not free money. Small-cap value stocks have higher expected returns because they are more sensitive to recessions, distress, and funding shocks; rational investors demand a premium to hold them. Behavioral economists like Lakonishok, Shleifer, and Vishny argued the opposite — that value works because investors systematically extrapolate too aggressively, making cheap stocks mispriced rather than riskier. The two camps imply very different futures: a risk premium should persist regardless of who knows about it; a behavioral mispricing should fade as it gets attention.

Which is true matters. If factors are risk premia, AI changes nothing about the long-run premium. If factors are behavioral, AI's main effect is to compress them faster, and the open question is whether new behavioral mistakes show up at the same rate the old ones get arbitraged out.

Why AI hasn't made the factors disappear

Three structural reasons keep the Fama-French factors alive in 2026, even with vast machine-readable data and millisecond execution:

The risk-premium component is unmovable. If small-value stocks really are riskier — more leveraged, more cyclical, more likely to disappear in a credit crunch — then no amount of AI can reduce that risk. The compensation has to be there, or no rational investor would hold them. AI can identify them faster, but it cannot change their underlying volatility, default rate, or correlation with the business cycle.

AI inherits human biases from its training data. Modern equity-market machine learning is trained on decades of price, fundamentals, and news data, all of which encode human overreaction, narrative chasing, and herd behavior. An AI trained on this corpus does not become a dispassionate Bayesian; it learns to predict the same patterns of overreaction it sees in the data. This is one reason momentum has not been arbitraged out — faster information processing has tightened up short-horizon momentum patterns and stretched longer-horizon mean reversion, but the underlying trend-following bias persists in the agents (human and machine) that generate the prices.

The capital that needs to arbitrage factors is constrained. Theoretical arbitrage assumes infinite, costless leverage. Real factor portfolios have meaningful tracking error against broad indexes, periods of multi-year underperformance (the HML factor had a 13-year drawdown from 2007 to 2020), and capacity limits that prevent any single fund from scaling indefinitely. Most institutional money is benchmarked against the S&P 500 and cannot tolerate the career risk of looking different for a decade. So the arbitrage doesn't fully complete.

Where the original factors have actually weakened

Honest factor analysis in 2026 has to admit that not all the original Fama-French factors have aged equally well.

Size, on its own, has been the weakest. Studies replicating SMB on post-1980 data have consistently found a much smaller premium than the original work suggested, and once you control for quality, the pure small-cap effect nearly vanishes. This is why the better small-cap factor products today screen for small and profitable and cheap, rather than just small.

Value went through what was probably its worst decade ever from 2010 to 2020, with growth stocks outperforming by such a margin that some researchers wondered whether the factor was structurally broken. The 2022 rate-shock partial reversion suggested it wasn't — but the multi-decade conviction needed to hold value through a 10+ year drought is the relevant filter, not whether the average historical premium is positive.

Momentum remains one of the most empirically robust factors but is also the most expensive to implement, with high turnover, an elevated tax-cost ratio, and brutal "momentum crashes" like March 2009 in which the strategy gives back several years of returns in weeks.

Quality and Profitability — the FF5 additions — have held up best in live, post-publication data, and they overlap most cleanly with what most long-horizon investors already want: cash-generative businesses with durable balance sheets.

Factor investing is a 20-year decision masquerading as a 6-month signal. The literature is strong on the long-run mean; the live experience is dominated by drawdown duration.

The factor ETF landscape: what the cost ladder looks like

The table below is a reference set of liquid factor ETFs an individual investor can actually buy in a US brokerage. Expense ratios and inception dates are pulled from each fund's most recent issuer fact sheet (linked below). No back-tested performance numbers are included here because a single regime — particularly the 2020-2025 window — is too short to be informative about long-run factor returns.

Source: issuer fact sheets — Avantis, Vanguard, Schwab Asset Management, iShares. Pulled 2026-05-05.

The cost spread is structural. A naive value tilt at 4 bp lives in a different cost regime than a smart-beta multi-factor wrapper at 35-50 bp. Over a 30-year horizon, a 30 bp annual fee gap on a $100,000 sleeve compounds to roughly $30,000 in foregone returns at 7% — a real, quantifiable drag that exists independently of whether the underlying factor produces any premium at all.

Macro frame: factor expectations in a 4.39% 10Y world

Factor returns are conditional on the macro regime. The 2010-2020 environment — zero rates, low inflation, persistent growth-multiple expansion — was structurally hostile to value and friendly to growth and long-duration momentum. The current regime, with the 10-year Treasury at 4.39%, the Fed funds rate at 3.64%, and CPI year-over-year at 3.3% (FRED, asof 2026-05-01 / 2026-04-01 / 2026-03-01), looks more like the 1990s than the 2010s.

Higher real rates compress the discount applied to long-duration cash flows, which mechanically hurts growth and helps value. Whether this regime persists for the next decade is unknowable, but it is a more historically typical backdrop for the factors than the prior decade was. The VIX at 16.99 sits below its long-run median, which has historically been a moderately favorable regime for momentum and quality and less favorable for low-volatility (which tends to lag in calm markets and outperform when stress arrives).

For deeper context on how a long-horizon investor should think about regime-dependent allocation, see The Full Framework for Scientific Investing, which lays out the broader methodology.

How a long-horizon investor should actually use this

Three guidelines come out of the empirical record:

1. Factor tilts are portfolio decisions, not trades. If you cannot stomach a 5- to 10-year underperformance window relative to a market-cap index, do not tilt. The expected reward depends on holding through the drought.
2. Cost discipline matters more than factor selection. The empirical premium of any single factor is small compared to the live-data dispersion across regimes. The fee you pay is certain; the premium is not.
3. Combine factors with independent drivers. Value and Momentum have historically been negatively correlated, so combining them reduces tracking error against the market without sacrificing much expected premium. This is the principle behind multi-factor products, though the implementation cost premium often eats the diversification benefit.

FAQ

1. Has AI made factor investing obsolete?
No, but it has compressed some short-horizon factor signals. The longer-duration factors that depend on either risk premia (value, profitability) or structural behavioral biases (momentum) remain visible in the data, with longer drawdown periods than the textbook backtests suggest.

2. Is the Size factor still real?
On a standalone basis, the post-1980 evidence for a pure size premium is weak. The size effect largely shows up only when small-cap is combined with quality and value screens, which is why modern small-value funds explicitly screen for profitability rather than buying the entire small-cap universe.

3. Why has Value underperformed for so long?
The 2010-2020 period combined zero interest rates, persistent growth-multiple expansion, and concentrated mega-cap leadership — three conditions that mechanically penalize value strategies. The premium reasserted itself partially in the 2022 rate shock. Whether that continues depends on whether real rates stay elevated.

4. Should I just hold the broad market and skip factors entirely?
For most individual investors, holding a low-cost broad-market index is a perfectly defensible choice. Factor tilting is an active decision that introduces tracking error against the benchmark, and the empirical premium is smaller than the variation across regimes. Skipping factors is not leaving money on the table; it is choosing a different risk profile.

5. How do AI-driven smart-beta ETFs compare to classic factor ETFs?
The marketing premise — that AI can dynamically allocate across factors better than a static weight — is plausible in principle and unproven in practice on a cost-adjusted basis. Most AI-branded ETFs charge 35-75 bp, which is a significant headwind versus a 4-15 bp single-factor product. For a deeper look at one such case, see QQQM vs. QRFT: Nasdaq 100 vs. AI-Driven Quality Factor Investing.

Key takeaways

• Factor returns survive AI for a structural reason: the risk-premium component cannot be arbitraged, and the behavioral component is reproduced rather than eliminated by machine-learning systems trained on human-generated data.
• Of the original Fama-French factors, Quality has aged best, Value has been real but with a brutal 2010-2020 drought, Momentum remains robust but expensive to implement, and standalone Size is the weakest.
• The expense ratio gap between low-cost factor ETFs (~6 bp) and smart-beta wrappers (~50 bp) is the most reliable, certain return driver an investor controls.
• Factor investing requires a 20+ year holding period. Drawdown duration, not the average historical premium, is the binding constraint for individual investors.
• The current 4.39% 10Y, 3.3% CPI macro regime is more historically typical than the 2010s and is a structurally less hostile backdrop for value and quality factors than the prior decade.

What this analysis can and can't tell you

This article frames the conceptual case for factor investing in 2026; it does not tell you which factor will outperform over the next five years. Live factor returns are noisy enough that a five-year window is essentially a single regime. The honest answer to "should I tilt to value now" is "if you are committed to holding through a potential 10-year underperformance window, yes; if you would abandon the position after three years of lag, no." Sample size, regime coverage, and the live-versus-backtest gap all argue for treating any factor tilt as a structural portfolio choice rather than a tactical decision.

Methodology

Macro context numbers (10Y Treasury, Fed funds rate, VIX, CPI YoY) are pulled from FRED with as-of dates noted in-text. ETF expense ratios and inception dates are taken from each issuer's most recent fact sheet, pulled 2026-05-05. No back-tested performance numbers are quoted because the readily available 5-year window represents a single macro regime and is uninformative about long-run factor returns. Factor literature references are to Fama & French (1992, 1993, 2015), Asness, Frazzini & Pedersen on quality, and Lakonishok, Shleifer & Vishny on the behavioral interpretation of value.

The editor maintains long-term ETF positions across broad-market and selected factor exposures; specific position sizes are not disclosed at the article level.

This article is for educational purposes and does not constitute personalized financial advice. See full Disclaimer.