Beyond the 3-Factor Model: Advanced Factor Investing

Extensions, Anomalies, and the Cutting Edge of Asset Pricing Research¶

Where This Fits¶

This is the fourth notebook in the series. The recommended reading order is:

1. Statistical Foundations — CLT, hypothesis testing, OLS, sandwich estimators, Newey-West derivation

2. Time Series Foundations — Stationarity, autocorrelation, volatility clustering, ergodicity

3. Fama-French 3-Factor Model — Data, regression mechanics, diagnostics, robust SEs, interpretation

4. Advanced Factor Models — this notebook

5. Practical Factor Investing — Real ETF analysis, portfolio construction, constrained optimisation

6. Solution Manual — Complete worked solutions to all exercises

Each notebook builds on the previous ones. The 3-Factor tutorial ended by asking: “can we do better with additional factors?” This notebook answers that question. The next notebook then applies these concepts to real-world factor ETFs.

What We’ll Cover¶

1. The Fama-French 5-Factor Model — Adding profitability (RMW) and investment (CMA) factors, and the dividend discount model foundation

2. The Carhart 4-Factor Model — Momentum (UMD/WML), fundamental momentum, and what you actually get in a momentum fund

3. Novy-Marx Gross Profitability — Why gross profitability predicts returns, the quality vs. profitability distinction, and the danger of multi-signal overfitting

4. Short-Term Reversals & Mean Reversion — Contrarian effects at different horizons

5. The Beta Anomaly & Betting Against Beta — Why CAPM beta fails empirically, and decomposing the low volatility anomaly

6. Meta-Research: What Survives Scrutiny? — Replication results, out-of-sample decay, transaction costs, the post-2005 efficiency shift, and the Adaptive Markets Hypothesis

7. The Great Debate — And What Practitioners Do About It — The case against and for factor investing, DFA/Avantis implementation, premium timing futility, skewness and diversification, and an honest verdict

8. Rolling-Window Analysis — Time-varying factor exposures and regime changes across market cycles

9. Summary and Further Reading — Key takeaways, the factor model hierarchy, and an annotated reading list

Prerequisites¶

• Statistical Foundations — especially the Newey-West HAC derivation (Section 7) and the CLT-to-p-value argument (Section 9). We use Newey-West t-statistics throughout this notebook; that tutorial explains why they are the right tool for time-series factor regressions.

• Time Series Foundations — especially stationarity (Section 2) and volatility clustering (Section 4). The rolling-window analysis in Section 8 of this notebook is motivated by the possibility that factor exposures are not stationary — a concept developed in that tutorial.

• Fama-French 3-Factor Tutorial — the core applied tutorial. We assume you are comfortable with OLS regression on factor data, interpreting alpha and factor loadings, and reading Newey-West standard errors from statsmodels output.

Quick Recap¶

In the earlier notebooks we built up a complete toolkit:

• Notebook 1 gave us the statistical machinery — from the CLT through OLS to Newey-West standard errors — all verified on hand-checkable examples.

• Notebook 2 showed why that machinery matters for financial data specifically: returns exhibit volatility clustering and weak autocorrelation, which is why HAC standard errors are the standard in empirical finance.

• Notebook 3 applied the toolkit to real data, regressing SPY, IWN, and VTV excess returns on Mkt-RF, SMB, and HML. We found that the market factor dominates SPY’s variation while SMB and HML capture IWN’s small-cap value tilt.

That 3-factor regression is our baseline throughout this notebook. Here, we ask: can additional factors — profitability, investment, momentum, reversals — improve on the 3-factor specification, and what does the answer tell us about the cross-section of expected returns?

Setup: Load Factor and ETF Data¶

Before diving into the theory and analysis, we need to assemble our dataset. We download:

1. Fama-French factor returns (3-factor, 5-factor, momentum, short-term and long-term reversal) from Kenneth French’s data library

2. ETF price data (SPY, IWN, VTV) from Yahoo Finance — the same three ETFs used in the 3-Factor Tutorial

The combined dataset spans 2000–2023, a longer sample than the introductory notebook. The additional years are deliberate: momentum crashes, value drawdowns, and other regime changes require a longer history to observe.

Section 1: The Fama-French 5-Factor Model¶

Key Definition: The Sharpe Ratio¶

Before we begin, one metric used frequently in this notebook is the Sharpe ratio, which measures risk-adjusted performance:

where $\bar{r}$ is the mean excess return (or factor return) and $\sigma_r$ is its standard deviation. A higher Sharpe ratio means better return per unit of risk. Typical values: 0.2–0.4 is moderate, >0.5 is strong, >1.0 is exceptional.

Note on inference: The Sharpe ratio is a descriptive statistic — it is just the sample mean divided by the sample standard deviation (both concepts from Section 1 of the Statistical Foundations notebook). A Sharpe ratio of 0.6 is not automatically “statistically significant.” To test whether a Sharpe ratio differs from zero, you need a t-test on the mean excess return — which is exactly what the regression intercept (alpha) and its Newey-West p-value provide. We derived Newey-West standard errors from first principles in Sections 7–9 of the Statistical Foundations notebook; here, we simply apply them.

To annualize from monthly data: multiply the monthly Sharpe ratio by $\sqrt{12}$ (equivalently, use annualized mean and annualized volatility).

A Note on the Sample Period¶

This notebook uses 2000–2023 data (vs. 2010–2023 in the introductory notebook). The longer sample is deliberate: the additional factors (especially momentum and reversals) exhibit dramatic regime changes (e.g., the 2009 momentum crash, value’s 2018–2020 drawdown) that a shorter window would miss. As we saw in the Time Series Foundations notebook, the properties of financial returns can vary across regimes — volatility clustering and structural breaks are real concerns. The longer sample lets us observe these dynamics directly.

From 3 Factors to 5¶

In 2015, Fama and French extended their original model by adding two new factors motivated by a growing body of evidence that profitability and investment patterns predict stock returns independently of size and value.

The Two New Factors¶

RMW — Robust Minus Weak (Profitability)

• Long stocks with robust (high) operating profitability, short stocks with weak (low) operating profitability

• Operating profitability = (Revenue − COGS − SGA − Interest Expense) / Book Equity

• Captures the empirical finding that profitable firms earn higher returns than unprofitable firms, controlling for other characteristics

• Average annual premium: roughly 3–4% historically (US data)

CMA — Conservative Minus Aggressive (Investment)

• Long stocks with conservative (low) asset growth, short stocks with aggressive (high) asset growth

• Investment = Change in total assets / lagged total assets

• Captures the finding that firms that invest aggressively tend to earn lower subsequent returns

• Average annual premium: roughly 2–3% historically

Why Add These Factors?¶

Fama and French (2015) presented three motivations for augmenting the original model:

1. Anomaly resolution. Many CAPM/3-factor anomalies (e.g., the profitability and asset-growth effects) disappear once you control for RMW and CMA.

2. Theoretical backing. A simple dividend discount model shows that — holding price constant — a stock’s expected return rises with expected profitability and falls with expected investment. This provides ex ante economic reasoning, not just empirical pattern-matching. We develop this derivation in full detail below — it is the theoretical heart of the five-factor model.

3. HML becomes redundant. In the 5-factor model, HML is largely spanned by the other four factors: its intercept (alpha) is near zero and statistically insignificant. This is one of the most debated results in modern asset pricing (see below).

The HML Redundancy Debate¶

A striking result from Fama-French (2015): when you include RMW and CMA alongside the market and size factors, the time-series variation in HML is almost fully explained, and HML’s alpha is close to zero.

Why? Value stocks tend to be:

• More profitable (positive RMW exposure)

• More conservative in their investment (positive CMA exposure)

So a portfolio that tilts toward profitability and low investment already captures much of what HML captures — the value premium may be a composite of profitability and investment effects.

This does NOT mean value investing is dead. It means:

• The economic channels through which value delivers returns are better understood: high book-to-market proxies for high expected profitability and conservative investment.

• Whether HML captures a distinct risk, or is merely a noisy proxy for RMW + CMA, remains actively debated. See Fama and French (2017) and Hou, Xue, and Zhang (2015) for opposing views.

The Dividend Discount Model Foundation of FF5¶

The theoretical cornerstone of the five-factor model — and what distinguishes it from a purely empirical exercise — is a dividend discount model (DDM) derivation presented in Section I of Fama & French (2015). This is worth developing carefully, because it provides the ex ante economic logic for why profitability, investment, and book-to-market should predict returns.

The Clean-Surplus Valuation Identity¶

Start from the standard present-value relation. The market value of a stock’s equity at time $t$ equals the discounted sum of expected future dividends:

where $M_t$ is the market value of equity, $d_{t+\tau}$ are dividends paid at $t+\tau$, $r$ is the long-run average expected return (the discount rate, assumed constant for simplicity), and $E_t(\cdot)$ denotes expectations at time $t$.

Now invoke the clean surplus accounting identity: all changes in book equity come either from earnings or dividends (and net share issuance, which we fold into dividends for simplicity):

where $B_t$ is book equity and $Y_{t+1}$ is total equity earnings in period $t+1$. Rearranging: $d_{t+1} = Y_{t+1} - \Delta B_{t+1}$, where $\Delta B_{t+1} = B_{t+1} - B_t$ is net investment in book equity.

Substituting into the DDM:

Dividing by Book Equity¶

Divide both sides by current book equity $B_t$:

This is the equation that generates the five-factor model. The left-hand side is the market-to-book ratio (the inverse of book-to-market). The right-hand side depends on three things: expected profitability ($Y/B$), expected investment ($\Delta B/B$), and the discount rate ($r$) — which is the expected return.

The Three Comparative Statics¶

Now hold different variables fixed and see what happens to $r$:

This is the key result. All three cross-sectional patterns — the value effect, the profitability effect, and the investment effect — follow from a single valuation equation. The DDM doesn’t tell us why discount rates vary across stocks (risk? mispricing?), but it tells us that if discount rates vary, these are exactly the observable characteristics that should predict returns.

Why This Matters¶

1. Theoretical coherence. The CAPM derives expected returns from a single risk factor. FF3 was largely empirical — Fama and French acknowledged in 1993 that their factors “may or may not be risk factors.” The DDM derivation gives FF5 a value-theoretic foundation: the factors are not arbitrary; they map onto the primitives of the valuation equation.

2. HML redundancy explained. The DDM shows that book-to-market captures a combination of profitability and investment information. Once you control for profitability (RMW) and investment (CMA) directly, book-to-market has little independent information left — explaining the famous HML redundancy result.

3. It applies to any theory of discount rates. Whether you think expected return variation is driven by rational risk compensation (efficient markets) or systematic mispricing (behavioral finance), the DDM comparative statics hold. The equation is an accounting identity combined with present-value logic — it’s agnostic about why $r$ varies.

4. Connection to the q-factor model. Hou, Xue, and Zhang’s (2015) q-factor model arrives at similar factors from production-based asset pricing (firms’ investment Euler equations). The fact that two very different theoretical starting points — demand-side valuation (DDM) and supply-side production theory — both point to profitability and investment as key state variables is powerful evidence that these factors capture something fundamental.

Fama & French (2015), p. 2: “The motivation for the five-factor model is the dividend discount model... If the model’s description of expected returns is complete, the intercept $a_i$ is zero for all assets.”

Limitations of the DDM Motivation¶

The derivation is not without caveats:

• Constant discount rates. The DDM derivation assumes a single constant $r$. In reality, discount rates are time-varying. Lettau and Wachter (2007) and others have explored how time-varying discount rates complicate the clean mapping from characteristics to expected returns.

• Clean surplus violations. In practice, accounting doesn’t perfectly satisfy the clean surplus relation (e.g., comprehensive income items bypass the income statement). This adds noise but doesn’t invalidate the qualitative predictions.

• The model is silent on momentum. The DDM naturally motivates value, profitability, and investment — but not momentum. This is one reason Fama and French excluded momentum from FF5: it lacks a clean valuation-theoretic justification, sitting uncomfortably between behavioral and risk-based explanations.

• Expectations vs. realizations. The DDM involves expected future profitability and investment. We observe only realized (past) values. The implicit assumption is that past profitability and investment are reasonable proxies for expectations — reasonable but imperfect.

Empirical Analysis: Factor Summary Statistics¶

With the theoretical foundation in place, let’s examine the empirical properties of the five factors. We’ll look at summary statistics, risk-return profiles, and the correlation structure — all of which inform how the factors interact in a regression.

HML–CMA Collinearity: Diagnostics and Practical Implications¶

The correlation matrix above reveals a strong positive relationship between HML (value) and CMA (conservative investment). This is not a coincidence — the DDM derivation in Section 1 shows that book-to-market is a composite signal that partially proxies for investment behaviour. Cheap stocks tend to be conservative investors, so HML and CMA carry overlapping information.

Why This Matters for Regressions¶

When two regressors are highly correlated, we have multicollinearity. This does not bias the OLS point estimates — $\hat{\beta}$ remains unbiased (Statistical Foundations, Section 5). But it inflates standard errors on the correlated coefficients, because the model struggles to distinguish each factor’s independent contribution. Concretely:

• Individual t-statistics on HML and CMA become unreliable — both may appear insignificant even if the joint explanatory power is strong

• Small changes in the sample or specification can flip the sign or magnitude of either coefficient

• The Variance Inflation Factor (VIF) quantifies this: $\text{VIF}_j = \frac{1}{1 - R_j^2}$, where $R_j^2$ is the R² from regressing factor $j$ on all other factors. A VIF above 5–10 is a conventional warning threshold

The Practitioner’s Solution: Drop CMA¶

Many practitioners and researchers (e.g., Blitz & Vidojevic, 2019; Fama & French themselves in robustness checks) note that when running a full 6-factor regression (FF5 + momentum), CMA is often the weakest and most redundant factor. Since HML already captures much of the investment effect (and CMA’s standalone premium is the smallest of the FF5 factors), a common approach is:

1. Run the full model with all factors — but check VIF and treat individual HML/CMA coefficients with caution

2. Drop CMA to obtain a 5-factor specification (Mkt-RF, SMB, HML, RMW, UMD) where HML absorbs the investment effect more cleanly

3. Compare the two specifications — if R² barely changes and alpha is stable, the dropped factor was redundant

This is not the same as saying CMA doesn’t matter economically. The investment effect is real (the DDM proves it). The issue is purely statistical: CMA and HML are too correlated to estimate their separate contributions reliably in the same regression.

Below we compute VIF scores and compare the full 6-factor regression with a 5-factor version that drops CMA.

Model Comparison: 3-Factor vs. 5-Factor¶

Now we fit both the original 3-factor model and the extended 5-factor model to our three ETFs. The key questions: does adding RMW and CMA improve the fit? Does the HML coefficient change — the “redundancy” effect Fama and French documented?

Interpreting the 3-Factor vs 5-Factor Comparison¶

Compare these results to the 3-factor regressions you ran in the 3-Factor Tutorial. The key question is: does adding RMW and CMA change the story?

What to look for:

1. R² improvement: The 5-factor model should explain at least as much variance as the 3-factor model. For broad-market ETFs like SPY, the improvement may be modest since the market factor dominates. For style ETFs like IWN (small value) or VTV (large value), the additional factors may capture meaningful variation.

2. Alpha changes: If alpha shrinks when moving from 3 to 5 factors, the “outperformance” was actually compensation for profitability/investment factor exposures — not genuine skill. Recall from the 3-Factor Tutorial that alpha should be interpreted carefully: a statistically insignificant alpha means the model fully explains the ETF’s returns.

3. RMW and CMA loadings:

   • IWN (small cap value) may show negative RMW — small value firms are often less profitable

   • VTV (large cap value) may show positive RMW — large value firms tend to be more profitable

   • Both value ETFs should show positive CMA — value stocks tend to be conservative investors

4. HML coefficient changes: Watch how β_HML changes between the 3- and 5-factor specifications. Due to the high correlation between HML and CMA, the HML loading may shrink when CMA is included.

Section 2: Momentum — The Carhart 4-Factor Model¶

The Premier Anomaly¶

Momentum is one of the most robust and persistent anomalies in finance. First documented formally by Jegadeesh and Titman (1993), it refers to the tendency of stocks that have performed well over the past 3–12 months to continue performing well, and vice versa.

The Carhart (1997) 4-Factor Model¶

Mark Carhart extended the Fama-French 3-factor model by adding a momentum factor:

UMD (Up Minus Down), also called WML (Winners Minus Losers):

• Long stocks with high returns over months $t-12$ to $t-2$

• Short stocks with low returns over the same period

• Month $t-1$ is skipped to avoid the short-term reversal effect (more on this later)

• Average annual premium: roughly 6–8% historically — larger than value or size

Why Does Momentum Exist?¶

This is one of the great debates in finance:

Behavioral Explanations:

• Underreaction: Investors are slow to incorporate new information → prices drift in the direction of news

• Overconfidence + Self-attribution bias: When investors’ picks go up, they attribute it to their skill and buy more; when picks go down, they blame bad luck

• Herding: Investors follow trends, creating self-reinforcing price movements

• Disposition effect: Investors sell winners too early and hold losers too long, creating persistent price trends

Risk-Based Explanations:

• Momentum stocks are riskier because they have crash risk — momentum strategies experience rare but devastating drawdowns (e.g., 2009 momentum crash: −73% in a few months)

• Time-varying risk: momentum profits may compensate for loading on a latent risk factor during economic state changes

• However, no fully satisfactory risk-based explanation has been widely accepted

Fama and French’s Position:

• Fama and French deliberately excluded momentum from their 5-factor model

• In their 2015 paper, they acknowledged momentum’s existence but argued it is not a “pervasive” risk factor in the same way as their other factors — it’s more of a trading strategy than a source of systematic risk

• They note that momentum does not have a clear risk-based story, which is uncomfortable from an efficient markets perspective

Fundamental Momentum: What’s Really Driving the Premium?¶

Robert Novy-Marx has argued (Rational Reminder, Episode 149) that much of what looks like price momentum is actually driven by a much older anomaly from the accounting literature: fundamental (earnings) momentum.

Earnings momentum (or post-earnings announcement drift, PEAD) was documented 20 years before price momentum. It refers to the tendency of stocks with improving year-over-year quarterly earnings to continue outperforming — there is a large initial price jump at the earnings announcement, but subsequent returns remain elevated for months afterward.

Novy-Marx’s decomposition reveals a critical asymmetry:

The implication: when you buy “price winners” (the standard momentum strategy), you are on average buying stocks that have seen earnings improvements — and it is the earnings improvement, not the price trend per se, that predicts future returns. Fundamental momentum is the more robust causal mechanism; price momentum is a noisy proxy for it.

What Are You Actually Getting in a Momentum Fund?¶

Novy-Marx identifies three distinct exposures embedded in price momentum strategies:

1. Earnings momentum — the genuine signal; stocks with improving fundamentals

2. Industry concentrations — industries tend to move together, so momentum portfolios become temporarily concentrated in hot sectors. This doesn’t help expected returns but increases volatility and contributes to occasional crashes

3. Time-varying market beta — in up markets, winners tend to be high-beta stocks (overweighting market risk); in down markets, relative winners tend to be low-beta stocks (underweighting market risk). This creates a perverse pattern: momentum is accidentally long the market just before downturns and short the market just before recoveries

These unintended bets help explain momentum’s notorious left-tail risk.

Why Novy-Marx is Skeptical of Momentum as a Standalone Factor¶

Despite momentum’s impressive historical average premium, Novy-Marx is cautious:

• High turnover: Academic momentum strategies have ~700% annual turnover. More cost-efficient implementations reduce the premium by roughly half.

• Crash risk: The 2009 crash (−50% in three months) and similar episodes in 1932 are not isolated flukes — they are structurally linked to how momentum concentrates in recently-winning betas and industries.

• Reduced diversification benefit once you include profitability: Momentum’s historical appeal was partly its negative correlation with value. But profitability already has a larger negative correlation with value. Once you trade value + profitability jointly, momentum’s incremental diversification benefit shrinks substantially — “it’s not a hedge anymore.”

• We don’t understand it: Unlike value and profitability, which have clear valuation-theoretic foundations (Section 1’s DDM derivation), momentum lacks a convincing risk-based explanation. This makes Novy-Marx “a little nervous” about its future persistence.

Novy-Marx suggests a more pragmatic use: embed momentum information into the implementation of other strategies (e.g., delay buying value stocks with negative short-term momentum, as DFA does) rather than running a standalone momentum portfolio. This captures some momentum benefit without incurring the full turnover and crash risk.

Momentum Crashes¶

A critical feature of momentum is its left tail risk. Daniel and Moskowitz (2016) documented that momentum experiences severe crashes, particularly after market downturns when previously losing stocks (now cheap) rapidly rebound:

• 2009: The momentum factor lost roughly 73% in a few months as value stocks rebounded after the financial crisis

• 1932: Similar crash after the Great Depression

• These crashes tend to occur when the market recovers sharply from a bear market — exactly when past losers (beaten-down stocks) surge

Section 3: Profitability — Novy-Marx and Beyond¶

Gross Profitability: The “Other Side of Value”¶

Robert Novy-Marx (2013) published one of the most influential papers in modern empirical asset pricing: “The Other Side of Value: The Gross Profitability Premium.” His key finding:

Gross profitability (Revenue − COGS) / Total Assets is the best single predictor of future stock returns among profitability measures, and its predictive power is comparable in magnitude to the value factor.

Why Gross Profitability?¶

Novy-Marx argued that gross profit (revenue minus cost of goods sold) is a cleaner measure of economic profitability than operating or net income because:

1. Less susceptible to accounting manipulation — Items further down the income statement (SGA, depreciation, special items) are more discretionary

2. Better captures “productive assets” — It measures the efficiency of a firm’s core operations before overhead and financing decisions

3. Controls for growth options — Firms with high gross profitability have valuable assets-in-place, and including lower-level expenses adds noise

As Novy-Marx explained on the Rational Reminder podcast (Episode 149):

“There’s a lot more information near the top of the income statement than at the bottom.”

The key intuition: people were “way too hung up on bottom line income.” Earlier researchers who tested profitability as a return predictor using net income found weak results — not because profitability doesn’t matter, but because lower-level accounting items add noise that obscures the signal.

The Value–Profitability Interaction¶

The most surprising result: gross profitability and value (book-to-market) are negatively correlated but both independently predict returns. This has profound implications:

• Controlling for value, profitable stocks earn higher returns

• Controlling for profitability, value stocks earn higher returns

• Combined value + profitability strategies significantly outperform either alone

• This is why Dimensional Fund Advisors (and later Avantis) emphasize both dimensions

Novy-Marx frames this through a Buffett lens: traditional Benjamin Graham value buys cheap stocks (low price relative to fundamentals), while profitability-based investing buys wonderful stocks at fair prices. Philosophically both strategies try to acquire productive capacity at a good price, but they look very different in practice — and crucially, they tend to outperform at different times. In U.S. large caps, value stocks tend to be less profitable, so a pure value strategy carries a “profitability headwind.” Trading value and profitability jointly removes that headwind and substantially improves the strategy’s reliability.

How Profitability Makes the Price Signal More Informative¶

The DDM tells us that a stock’s price reflects both expected future cash flows and the discount rate (expected return). A low price can mean either a high discount rate (what we want: a cheap stock with high expected returns) or low expected cash flows (a value trap).

Profitability resolves this ambiguity. Among stocks with similar profitability, price differences are more likely driven by discount rate differences — which is exactly the variation that predicts returns. Novy-Marx’s insight is that profitability doesn’t independently identify high-return stocks so much as it makes the value signal cleaner: “Once you condition on profitability, on average, when you buy a low price firm, you’re buying a firm that has a higher discount rate.”

This is the deeper explanation for why value + profitability outperforms either alone.

Fama-French RMW vs. Novy-Marx GP/A¶

The Fama-French RMW factor uses operating profitability (a broader measure), while Novy-Marx advocated for gross profitability (GP/A). The two are related but not identical:

Both capture the profitability premium, but the exact measure matters. Recent research (Ball et al., 2015, 2016) showed that operating profitability may actually be slightly better than gross profitability when defined relative to book equity — partly reconciling the two approaches.

Profitability vs. “Quality” — A Critical Distinction¶

When Novy-Marx began presenting his profitability research to practitioners, the most common response was: “Oh yeah, we already run a quality strategy.” But each person described something different. This conflation is a persistent source of confusion in factor investing.

Novy-Marx’s critique (Rational Reminder, Episode 149):

“For me, quality is really... I think most people would call profitability a quality strategy, but the problem for me with quality is that it’s not a well-defined thing that means the same thing to everyone. It’s sort of this general thing that I kind of view as a marketing term, more than something that’s really meaningful about investment.”

The term “quality” gained traction in the early 2000s after massive withdrawals from growth funds during the NASDAQ deflation. It was, in Novy-Marx’s view, a way to repackage growth-like strategies for investors burned by the tech bust — “people who just have a growth fund bend.” Quality products typically combine multiple signals — low earnings variability, high return on equity, low leverage, and others — into a composite score.

The multi-signal overfitting problem: This composite approach is where Novy-Marx sees the real danger. His formal critique (Novy-Marx, 2015, “Backtesting Strategies Based on Multiple Signals”) showed that combining multiple signals into a single score can produce unreasonably strong backtests even when the individual signals are weak or statistically unreliable:

1. Pick 20 candidate quality measures

2. At the 5% significance level, on average one will look significant by chance alone

3. Combine the best three into a Z-score composite

4. The composite will look extremely reliable — not because the signals work, but because diversifying across signal noise reduces tracking error, making the backtest’s t-statistic appear inflated

Novy-Marx uses a vivid analogy: “It’s as if I let 20 monkeys throw darts at the Wall Street Journal, and none of them is likely to have lights-out performance, but if I pick the best three and I diversify across their recommendations, it’s going to look great.”

How to evaluate multi-signal strategies:

• Demand evidence that each individual signal contributes meaningfully to performance on its own

• If a strategy uses seven signals and four are individually insignificant, those four may be inflating the backtest without adding real predictive power

• Signals that are negatively correlated with your existing strategy can be valuable (diversification), but they still need to individually justify their inclusion

• Be skeptical of any strategy whose backtested performance looks “too good” — the more signals combined, the easier it is to engineer impressive-looking results

“These premia that we talk about in these factor models are all fairly weak in some sense and they have to be. I don’t believe that there’s Sharpe ratios of three lying around for everyone to exploit.”

Key takeaway: Profitability is a well-defined, theoretically grounded factor with a clear economic interpretation. “Quality” is an umbrella term that may or may not include profitability and often bundles in signals that individually lack statistical support. For factor investors, the distinction matters: use profitability, be wary of “quality.”

Profitability: Risk or Mispricing?¶

Risk-based view (Fama-French perspective):

• Profitable firms are riskier because their cash flows are more sensitive to economic conditions

• High profitability implies high expected returns as compensation for this risk

Mispricing view:

• Investors undervalue profitable firms because they focus too much on price multiples

• The market is slow to recognize that high profitability is persistent

Novy-Marx’s view: The risk-based intuition for profitability is counterintuitive — more profitable firms seem safer, not riskier. As he explained on Rational Reminder (Episode 149): “You’d think that the more profitable firms... might be more stable and they might provide some insurance against bad times.” His resolution is to sidestep the risk debate entirely: the DDM identity requires that, holding price constant, higher profitability implies higher expected returns — regardless of whether that expected return is compensation for risk or a result of mispricing. The valuation equation holds as an accounting identity; it doesn’t depend on the mechanism.

The profitability premium is also “too clean” to be mispricing — it works in every size group, in international markets, and persists after transaction costs. This favors a risk-based view, though the debate continues.

Section 4: Short-Term Reversals, Long-Term Reversals, and Mean Reversion¶

The Time Horizon Puzzle¶

One of the most fascinating patterns in asset pricing is that the sign of return predictability flips depending on the horizon:

This creates a rich “U-shape” of return predictability across time.

Short-Term Reversals (1-Month)¶

Documented by Jegadeesh (1990) and Lehmann (1990):

• Stocks that had high returns last month tend to have low returns this month, and vice versa

• This is why the standard momentum factor skips month $t-1$ — it would contaminate the signal

• Explanation: Primarily driven by liquidity provision and microstructure effects:

  • Temporary price pressure from large trades

  • Bid-ask bounce

  • Market-maker inventory management

  • Overreaction to short-term news

Short-term reversal strategies have high gross returns but are extremely expensive to trade — they require monthly rebalancing with very high turnover. After transaction costs, much of the premium disappears for most investors.

Long-Term Reversals (3–5 Years)¶

Documented by DeBondt and Thaler (1985):

• Stocks with poor returns over the past 3–5 years tend to outperform over the next 3–5 years

• This is closely related to the value effect — stocks that have declined for years become cheap (high book-to-market), and cheap stocks eventually recover

• Explanation: Overreaction hypothesis — investors extrapolate past trends too far. After years of bad performance, expectations become overly pessimistic, creating buying opportunities.

Mean Reversion in Financial Markets¶

Mean reversion is the broader concept that extreme values tend to revert toward the average over time. In finance:

Asset-Level Mean Reversion:

• Individual stock returns mean-revert over long horizons (DeBondt & Thaler, 1985)

• But NOT over short horizons (1–12 months) — this is where momentum works

• Valuation ratios (P/E, P/B) mean-revert as well — when valuations are extreme, subsequent returns tend to push them back toward historical norms

Factor-Level Mean Reversion:

• Factor premia themselves mean-revert — after periods of poor performance, factors tend to recover

• This is particularly relevant for the value factor, which experienced a historic drawdown from ~2018–2020 and subsequently recovered

• Investors who understand this can maintain conviction during drawdowns

Market-Level Mean Reversion:

• Aggregate stock market returns show weak mean-reversion over long horizons (Fama & French, 1988)

• High valuations (CAPE ratio) predict lower subsequent long-run returns

• But short-run market returns are nearly unpredictable

Connection to the Fama-French Framework¶

The reversal factors connect to the broader factor model:

• Long-term reversals ≈ Value (HML): Buying long-term losers is essentially buying cheap stocks

• Short-term reversals ≠ Momentum: They work in opposite directions and at different horizons

• The momentum factor skips month $t-1$ precisely to avoid contaminating momentum with the short-term reversal signal

Section 5: The Beta Anomaly and Betting Against Beta¶

The CAPM’s Most Embarrassing Prediction¶

The CAPM makes a clean, testable prediction: expected excess returns are proportional to market beta.

Higher beta → higher expected return. Period. This is the theoretical foundation of much investment advice: take more risk, earn more return. It underpins the Bogleheads philosophy (“hold the whole market”), target-date fund glide paths, and the basic intuition that equities outperform bonds because they’re riskier.

The problem: empirically, the security market line is flat — or even negatively sloped.

This has been known since the earliest cross-sectional tests:

• Black, Jensen & Scholes (1972): Sorted stocks into beta deciles. Found that low-beta stocks earn more than the CAPM predicts, and high-beta stocks earn less. The realized SML is far flatter than the theoretical line.

• Fama & MacBeth (1973): Pioneered the cross-sectional regression methodology that became the standard. Confirmed the flat SML.

• Frazzini & Pedersen (2014): Formalized this as the “Betting Against Beta” (BAB) factor — long leveraged low-beta stocks, short de-leveraged high-beta stocks — and showed it earns significant risk-adjusted returns across countries and asset classes.

• Bali, Engle & Murray (2016): In Empirical Asset Pricing: The Cross Section of Stock Returns, they provide a thorough textbook treatment showing that market beta has essentially zero explanatory power for the cross-section of expected returns, once other characteristics are controlled for.

Why Does This Happen?¶

Several explanations have been proposed:

1. Leverage Constraints (Frazzini & Pedersen, 2014) The most influential explanation. Many investors — mutual funds, pension funds, retail investors — face constraints on leveraging. If you want more return than the market but can’t use leverage, you must buy high-beta stocks. This excess demand bids up high-beta stocks and depresses their future returns, while low-beta stocks are shunned and become underpriced.

2. Lottery Demand (Bali, Cakici & Whitelaw, 2011) Some investors are attracted to stocks with lottery-like payoff distributions — high volatility, high skewness, low probability of a huge payoff. High-beta stocks tend to have these characteristics. This “lottery demand” pushes up their prices and depresses expected returns.

3. Benchmarking and Tracking Error (Baker, Bradley & Wurgler, 2011) Professional fund managers are evaluated relative to benchmarks (e.g., the S&P 500). Buying low-beta stocks creates tracking error — the portfolio deviates significantly from the benchmark. Career-concerned managers avoid this, creating systematic demand for high-beta stocks and neglect of low-beta stocks.

4. Disagreement and Short-Sale Constraints (Hong & Sraer, 2016) High-beta stocks amplify disagreement among investors (optimists and pessimists disagree more when the stock is more sensitive to the market). Combined with short-sale constraints, this causes high-beta stocks to be overpriced (pessimists can’t short them), which lowers their subsequent returns.

The BAB Factor Construction¶

Frazzini and Pedersen construct the BAB factor as follows:

1. Rank all stocks by estimated beta

2. Form two portfolios: low-beta (below median) and high-beta (above median)

3. Leverage the low-beta portfolio up to beta = 1 and de-leverage the high-beta portfolio down to beta = 1

4. BAB = Leveraged low-beta portfolio − De-leveraged high-beta portfolio

Both legs have beta ≈ 1, so the spread is approximately market-neutral. Any non-zero return is pure alpha.

The Critique: Does BAB Survive Implementation?¶

Not everyone is convinced:

• Novy-Marx & Velikov (2022), “Betting Against Betting Against Beta”: BAB profits are heavily concentrated in micro-cap stocks that are expensive or impossible to trade. After realistic trading costs and excluding micro-caps, BAB returns shrink dramatically.

• Cederburg & O’Doherty (2016), “Does It Pay to Bet Against Beta?”: Show that conditional (time-varying) beta models substantially reduce the BAB premium.

• Liu, Stambaugh & Yuan (2018): Argue beta-related anomalies are more about idiosyncratic volatility than beta per se.

Decomposing the Low Volatility Anomaly: It’s Really About Small Unprofitable Growth¶

The “low volatility anomaly” — often marketed as a distinct defensive strategy — gets a thorough decomposition from Novy-Marx (Rational Reminder, Episode 149). His analysis reveals that the anomaly is largely a repackaging of known factor exposures, not an independent phenomenon.

Step 1: What characterizes the highest-volatility stocks?

Novy-Marx asks a simple diagnostic question: if you had to predict which stocks are most volatile without looking at volatility directly, what characteristics would you choose?

Step 2: The interaction effect

Average growth stocks and average value stocks have roughly the same volatility. But growth stocks are typically more profitable. Since profitability is associated with lower volatility, the equal-volatility observation implies that high relative price (growth) is associated with higher volatility — once you control for profitability. The most volatile stocks sit in the intersection: small, unprofitable, expensive (growth).

Step 3: This corner of the market has dramatically underperformed

The small unprofitable growth segment has been “by far the worst underperforming segment of the market” in the post-1968 sample. This underperformance fully explains the “low volatility effect”:

“The high volatility stocks unquestionably have performed really poorly in the post-'68 sample, but I think they do so primarily because small unprofitable growth stocks, which tend to be high volatility, have been by far the worst underperforming segment of the market.”

Step 4: Low-beta strategies and factor loadings

Novy-Marx extends the analysis to beta arbitrage strategies (Betting Against Beta). Low-beta portfolios carry large tilts toward:

• High profitability (positive RMW loading)

• Conservative investment (positive CMA loading)

• A modest value tilt — but this value tilt is really proxying for the investment loading

Once you control for profitability and investment in the five-factor model, the alpha on low-beta strategies shrinks substantially. The “BAB premium” is largely compensation for factor exposures that DFA and Avantis already capture through their integrated approaches.

Implications for “low-vol” products:

• A simple high-volatility exclusion (removing the most volatile 1–2% of stocks by market cap) is reasonable — it avoids the worst-performing corner of the market at negligible cost.

• A dedicated “low-volatility” product is more problematic: it’s an inefficient way to get what is essentially a small-unprofitable-growth exclusion combined with profitability and investment tilts. If you already hold value + profitability factor portfolios, a low-vol fund adds little incremental value and may create unintended concentration.

• Novy-Marx: “I kind of view the low volatility or low beta in the same way [as quality] — again, I do think that what they’re doing is getting you these other tilts.”

What Does This Mean for Bogleheads?¶

The beta anomaly is often cited as contradicting the simple “hold the market” advice. But the implications are nuanced:

1. The market portfolio is still efficient in the Fama-French sense — it captures the equity risk premium, which is large and robust. The anomaly is about the cross-section (low-beta vs. high-beta) within equities, not about whether equities beat bonds.

2. Implementation matters: As Novy-Marx and Velikov showed, actually capturing the BAB premium after costs is much harder than the academic long-short factor suggests. Most individual investors cannot efficiently leverage low-beta portfolios.

3. Factor tilts subsume much of it: Stocks that are low-beta also tend to have high profitability and low investment. Much of the “low-beta alpha” is captured by RMW and CMA tilts — exactly what DFA and Avantis already do. The value tilt (HML) plays a negligible role once profitability and investment are controlled for.

4. The deeper lesson: The CAPM is not a complete model. The FF 5-factor model (and its extensions) does a much better job of pricing the cross-section. The flat SML is an indictment of the CAPM’s beta-only view, not of factor investing generally.

Empirical Demonstration¶

Below we download beta-sorted decile portfolio returns from Kenneth French’s data library and visualize the relationship between beta and average returns — the empirical security market line.

Section 6: Meta-Research — What Survives Scrutiny?¶

The individual factors we studied in Sections 1–5 all have impressive in-sample track records. But a body of research — primarily from Andrew Chen (Federal Reserve Board), Tom Zimmermann, Mihail Velikov, and Robert Novy-Marx — asks harder questions: can you replicate these findings? Do they hold up out of sample? And can anyone actually capture the premiums after trading costs?

This section synthesizes the meta-research. The conclusions are more nuanced than either the factor-bullish or factor-skeptical camps typically admit.

6.1 The Replication Landscape¶

Chen & Zimmermann, “Open Source Cross-Sectional Asset Pricing” (2022) Published in Critical Finance Review | SSRN 3604626

Chen and Zimmermann replicated over 200 published cross-sectional return predictors using standardized methodology, publishing all code and data at openassetpricing.com. Key findings:

• Of ~300 variables examined, roughly 200 were actually shown to predict returns in the original papers (many widely-cited “anomalies” never had significant results in the first place)

• Of those 200, Chen & Zimmermann successfully replicated all but three — a ~1–2% failure rate

• This contradicts the popular “replication crisis” narrative from Hou, Xue & Zhang (2020), who reported widespread failure. Chen argues that paper conflated confirmed null results with failed replications — counting cases where the original paper itself found no significance as “replication failures”

• The estimated false discovery rate in cross-sectional asset pricing is less than 10% — far lower than the alarming claims in Harvey, Liu & Zhu (2016), who proclaimed “most claimed research findings in financial economics are likely false.” Chen dismantles that claim: it conflated insignificant with false (a statistical error) and many with most (a rhetorical overreach)

The bottom line: The published findings in cross-sectional asset pricing are reliable. There is no replication crisis. The anomalies are real statistical regularities, not fabrications. The strongest, most replicable factors remain: market, value, momentum, profitability, investment, and short-term reversal.

6.2 Out-of-Sample Decay and Transaction Costs¶

The anomalies are real — but can you profit from them?

Out-of-sample decay (McLean & Pontiff, 2016; Chen & Zimmermann, 2022):

Post-publication, anomaly returns decline by roughly 50%. The decline is gradual — in the first few years, only ~25%. If anomalies were pure statistical noise, returns would vanish instantly upon replication. Instead, the gradual decay is consistent with real but transient mispricing that gets arbitraged away.

Transaction costs (Chen & Velikov, “Zeroing in on the Expected Returns of Anomalies,” 2023): Published in the Journal of Finance | SSRN 3073681

Chen and Velikov combine out-of-sample decay with trading costs (measured as effective bid-ask spreads):

• Transaction costs consume roughly 25–30% of anomaly returns in the original sample periods

• Combined with the ~50% out-of-sample decay, the residual is close to zero

• After the post-2005 efficiency gains (Section 6.4), even the best individual anomalies earn effectively nothing net of basic transaction costs

As Chen stated on Rational Reminder (Episode 316, August 2024):

“Out-of-sample, the returns are basically completely gone after trading cost... even the best of these anomalies or predictors earns nothing in this very baseline transaction cost case.”

Transaction cost caveats (which Chen himself notes):

• Costs assume aggressive market orders for a small trader — no price impact modelling

• Short-sale costs are ignored (which would make things worse)

• Sophisticated cost-mitigation techniques are not fully modeled (which could help)

• A separate paper finds that short-sale costs alone can eliminate anomaly profits

Novy-Marx & Velikov — Which factors survive implementation?

Novy-Marx and Velikov have systematically evaluated factor implementability across several papers:

• “A Taxonomy of Anomalies and Their Trading Costs” (2016): Many anomalies that appear profitable in academic backtests disappear after trading costs. High-turnover strategies (like short-term reversal) are particularly vulnerable; low-turnover strategies (value, profitability, investment) survive much better.

• “Betting Against Betting Against Beta” (2022): BAB profits are largely concentrated in difficult-to-trade micro-cap stocks. After realistic costs and size exclusions, BAB profits shrink dramatically.

• “Assaying Anomalies” (2023): Systematic evaluation of which anomalies survive: value, profitability, and momentum are robust survivors. Many exotic anomalies and signals concentrated in micro-caps are not.

The key distinction: Chen is right that most of the “factor zoo” is worthless after costs. But the core factors that sophisticated implementers like DFA and Avantis tilt toward are precisely the ones Novy-Marx & Velikov identify as survivors.

6.3 Does Peer-Reviewed Theory Help Pick Factors?¶

Chen, Lopez-Lira & Zimmermann, “Does Peer-Reviewed Research Help Predict Stock Returns?” (2024) Working paper | SSRN 4308069

Perhaps the most philosophically unsettling finding in Chen’s research program. He categorizes published anomalies by theoretical underpinning — mispricing story, risk-based story, toy model, or full quantitative equilibrium model — and tests whether sophisticated theory helps:

• Factors backed by strong theoretical models perform no better out of sample than naively data-mined accounting ratios

• If anything, the more rigorous the theoretical justification, the worse the out-of-sample performance

• Purely mechanical data mining (searching through 29,000 accounting ratios) could have uncovered all major anomalies decades before they were published

“If you could data mine [accounting ratios] in 1980, you could have uncovered all of these anomalies long before they were published, even decades before.”

This challenges the widely held belief that economic theory distinguishes “real” factors from data-mined artifacts. The theoretical apparatus behind our DDM derivation in Section 1 — and the risk-based stories for value, profitability, and investment — may add less informational value than commonly assumed.

6.4 The Post-2005 Structural Break¶

A key empirical regularity across multiple studies: factor performance exhibits a “kink” around the mid-2000s. Returns are substantially weaker from 2005 onward.

Two competing hypotheses:

1. Technology and efficiency: The internet and IT revolution made information dramatically cheaper and faster to process. In the 1963–1989 era (the data behind Fama-French 1993), trading on book-to-market required receiving accounting statements by mail and making phone calls. Now it takes microseconds.

2. Publication effect: Most major anomaly papers were published around 2000–2005, and subsequent arbitrage traded the premia away.

Chen favors the technology story, though both forces likely reinforce each other. Either way: historical factor premia reflect a less efficient informational environment that no longer exists. Using 1963–2005 average returns to estimate future premia likely overstates them.

6.5 Markets Adapt: The Adaptive Markets Hypothesis¶

Chen frames his findings through Andrew Lo’s Adaptive Markets Hypothesis — perhaps the most important theoretical lens for understanding these results:

• Markets are not perpetually efficient (as classical EMH suggests) but become more efficient over time as participants learn and compete

• Mispricings are real when they first appear, but they get corrected — through academic publication, technology, market evolution, or all three

• Factor premia are therefore transient mispricing, not permanent risk compensation

• The traditional efficient markets view assumed markets are efficient all the time, which paradoxically led researchers to explain away mispricings with risk theories (creating “factors”). Chen suggests these risk-based stories were motivated reasoning

“These are all mispricings that are transient. If you’re a practitioner, you’ve got to get to something new.”

This is a profound shift from the traditional factor investing playbook. If premia are adaptive rather than structural, “buy and hold factor tilts forever” may be overly optimistic.

6.6 Alternative Factor Frameworks¶

The q-Factor Model (Hou, Xue, and Zhang, 2015, 2021)

An alternative to the Fama-French framework from production-based asset pricing theory:

• $r_{ME}$: Size factor (similar to SMB)

• $r_{ROE}$: Profitability factor based on return on equity (similar to RMW)

• $r_{I/A}$: Investment factor (similar to CMA)

The q-factor model is motivated by firms’ investment Euler equations (supply-side) rather than the consumption-based (demand-side) arguments of the traditional CAPM. The fact that two very different theoretical starting points — demand-side valuation (DDM) and supply-side production theory — both point to profitability and investment as key state variables is powerful evidence that these factors capture something fundamental. Empirically, the q-factor model has similar performance to FF5 and subsumes many anomalies.

Asness (AQR) — Value and Momentum Everywhere

Cliff Asness, Tobias Moskowitz, and Lasse Pedersen (2013) showed that value and momentum premia exist across asset classes — not just equities but also bonds, currencies, and commodities. This cross-asset evidence is strong evidence against US-specific data mining.

Asness has also been a prominent voice during the value factor’s worst-ever drawdown (2018–2020), arguing it was driven by extreme “momentum-fueled growth speculation” rather than a fundamental change in the value premium. That prediction proved correct: value recovered dramatically in 2021–2022. Current (2025–2026) value spreads remain wide by historical standards, suggesting above-average expected value premia going forward.

Empirical Illustration¶

The code below shows factor performance by decade and the comprehensive 6-factor model — illustrating the time-variation in premia that this section’s research documents.

Key references:

• Chen, A. & Zimmermann, T., “Open Source Cross-Sectional Asset Pricing” (Critical Finance Review, 2022). SSRN

• Chen, A. & Velikov, M., “Zeroing in on the Expected Returns of Anomalies” (Journal of Finance, 2023). SSRN

• Chen, A., Lopez-Lira, A. & Zimmermann, T., “Does Peer-Reviewed Research Help Predict Stock Returns?” (2024). SSRN

• McLean, R.D. & Pontiff, J., “Does Academic Research Destroy Stock Return Predictability?” (Journal of Finance, 2016). SSRN

• Harvey, C.R., Liu, Y. & Zhu, C., “…and the Cross-Section of Expected Returns” (Review of Financial Studies, 2016). SSRN

• Novy-Marx, R. & Velikov, M., “Assaying Anomalies” (2023)

• Novy-Marx, R. & Velikov, M., “A Taxonomy of Anomalies and Their Trading Costs” (2016)

• Hou, K., Xue, C. & Zhang, L., “Replicating Anomalies” (Review of Financial Studies, 2020). SSRN

• Asness, C., Moskowitz, T. & Pedersen, L., “Value and Momentum Everywhere” (Journal of Finance, 2013)

• Lo, A., The Adaptive Markets Hypothesis (2017)

Section 7: The Great Debate — And What Practitioners Do About It¶

Section 6 presents a sobering picture: individual anomalies largely vanish after costs in the modern era, theory does not predict which factors will persist, and markets appear to be adapting. This naturally fuels the longstanding debate between total-market indexers (“Bogleheads”) and factor-tilted investors. Here we lay out both sides honestly, then examine what sophisticated practitioners actually do.

7.1 The Case Against Factor Investing¶

Andrew Chen’s research provides the strongest evidence for the Boglehead position:

“Is everything I was taught about cross-sectional asset pricing wrong?!” — Title of Rational Reminder Episode 316 (August 2024)

7.2 The Case For: Why Implementation Is Not Replication¶

Chen’s analysis focuses on individual anomalies traded as academic long-short portfolios. But nobody actually invests this way. Several counter-arguments qualify his strongest conclusions:

Multi-factor integration ≠ individual anomaly trading. Chen himself acknowledges this: “Transaction costs are actually, in a sense, lower if you use more predictors at the same time.” When one signal calls for a buy and another calls for a sell, the trades offset — reducing turnover. DeMiguel, Martín-Utrera, Nogales & Uppal (Review of Financial Studies, 2020) confirm that optimally combining predictors while accounting for transaction costs yields portfolios that survive real-world costs.

Long-only ≠ long-short. Academic factor returns assume both a long side (buy winners) and a short side (short losers). The short side is expensive (borrowing costs, recall risk, short-sale constraints) and often concentrated in micro-caps. Real-world factor investors like DFA and Avantis use long-only tilts — overweighting stocks with favorable characteristics, underweighting those without — which is dramatically cheaper. As Gerard O’Reilly (DFA’s co-CEO and CIO) discussed on Rational Reminder Episode 360 (June 2025), DFA’s implementation operates at a fundamentally different cost level than academic long-short replication.

The core factors survive. Novy-Marx & Velikov’s “Assaying Anomalies” (2023) specifically identifies value, profitability, and momentum as factors that survive realistic implementation costs. Chen’s “the whole factor zoo is zero” applies to the zoo, not necessarily to the handful of factors at the zoo’s core.

The DDM identity holds regardless. The dividend discount model derivation from Section 1 guarantees that — holding cash flows constant — cheaper stocks have higher expected returns. This is an accounting relationship, not a market efficiency claim. It doesn’t depend on whether premia represent “risk” or “mispricing.”

Risk vs. mispricing is not settled. If value stocks genuinely compensate investors for bearing distress risk, that compensation should persist regardless of market efficiency improvements. Chen’s finding that risk-based factors perform no better out of sample weakens this view — but the dramatic value recovery of 2021–2022 after the deepest drawdown in history is hard to reconcile with “the premium is permanently gone.”

Value spreads are informative and currently wide. Wei Dai at Dimensional has shown that the current gap between cheap and expensive stocks predicts future value premia in the cross-section. As of 2025–2026, value spreads remain wide, suggesting above-average expected value premia going forward.

Avantis’ “expected return” approach (Repetto, Rational Reminder Episode 315, July 2024): Avantis ranks stocks by a composite expected-return measure combining valuation and cash-flow profitability. This is more similar to Chen’s “data mining” approach than to traditional factor replication — if Chen is right that data mining works as well as theory, this may actually validate Avantis’ approach.

7.3 Wei Dai and Dimensional: From Theory to Portfolio¶

Wei Dai, Head of Investment Research at Dimensional Fund Advisors, has published extensively on how to translate academic factor research into real portfolios. Her work — discussed in depth on Rational Reminder (Episode 306, May 2024) — demonstrates that the how of factor implementation matters as much as the what.

Premiums for the Long Run (Dai, 2022):

• Even with statistically significant premia, investors often need 10–15+ years to have a high probability (>80%) of outperforming the market

• Factor investing requires patience — multi-year tracking error vs. the market is normal, not a failure

• The probability of outperforming increases with horizon, but meaningful multi-year stretches of underperformance are expected

The Integrated Approach

Dai identifies three common approaches to multi-premium portfolios:

1. Combination: Separate single-factor portfolios (e.g., ⅓ small-cap, ⅓ value, ⅓ high-profitability), then merge. Problem: interactions create offsetting tilts — small-cap deep value stocks tend to be low-profitability, so the combination yields uncontrolled exposures.

2. Market-Satellite: Core market exposure plus a concentrated factor satellite. Problem: you must “push harder” on a tiny corner of the market, creating extreme overweights and high trading costs in the least liquid segment.

3. Integrated (DFA’s approach): Start with the market and do a three-way sort on size, value, and profitability simultaneously. Gradually shift weight from segments with lower expected returns to those with higher expected returns — like “tilting an ice cube tray.”

The integrated approach dominates on all three evaluation dimensions: expected return, risk control, and cost efficiency.

The Multiplier Weighting Scheme

Within the integrated approach, DFA uses a price-based “multiplier” system:

• Stocks are sorted into groups by their intersection on three dimensions (market cap × relative price × profitability)

• Each group is weighted by its total market capitalization × a multiplier (e.g., 2× for high-expected-return segments, 0.5× for low)

• Within each group, stocks are held in proportion to their market cap weights

This avoids the extreme deviations of non-price-based schemes, which can overweight micro-caps by 50–100× their market weight while holding mega-caps at basis points. Those deviations produce “shiny performance on paper” that cannot survive real-world trading.

Short-Term Reversals as an Implementation Tool

In collaborative research with Novy-Marx (Financial Analysts Journal, 2024), Dai shows that short-term price reversals are driven by liquidity provision. Their innovation: clean the reversal signal by adjusting for post-earnings announcement drift and short-run industry momentum, then use it as a reason not to trade — delay buying recent winners and delay selling recent losers until the reversal dissipates. Estimated value-add: 1–10 basis points without additional turnover — small but meaningful when “you’re fighting for every basis point.”

Don’t combine value and profitability in separate funds. Novy-Marx emphasizes (Rational Reminder, Episode 149) that buying a value ETF and a profitability ETF separately can cancel out factor exposures (because value stocks tend to be less profitable and vice versa). Integrated approaches — where a single portfolio jointly sorts on value and profitability — are substantially more efficient. This is the approach both DFA and Avantis use.

Selecting a Systematic Manager

Knowing the academic factors is not an investment solution. Key criteria:

1. Rigorous research with proper incentives to avoid overfitting

2. Implementation expertise — portfolio design, daily rebalancing infrastructure, technology + human oversight

3. Look beyond performance — examine periods of underperformance (did the manager maintain exposure and capture premiums when they showed up?)

4. Fee structure — performance fees incentivize increased tracking error; fixed fees better align with the patient, consistent focus that systematic strategies require

7.4 Premium Timing Is (Almost Certainly) Futile¶

DFA tested 720 premium timing strategies combining 3 signals × 5 switching thresholds × 2 rebalancing frequencies × 2 historical distribution methods × 4 premiums × 3 regions. Results: only 30 out of 720 (4.2%) “worked” — and even the best was extremely fragile:

• The top performer (timing US market premium using valuation ratios) generated 5.5% excess return and avoided the 2001/2008 downturns

• But changing any single parameter — the percentile threshold, the rebalancing frequency, or the rolling window length — cut excess return by more than half

• The same strategy didn’t work in other regions or for other premiums

• The analysis didn’t even account for taxes or trading costs

“If your results are driven by some outliers, a very small part of the sample, then you need to ask yourself how confident you are to put your money into a strategy that has worked well in historical data but is actually very sensitive to just a few outliers.”

Novy-Marx’s Satirical Demolition

Novy-Marx takes the timing critique further. In his satirical paper (described on Rational Reminder, Episode 149), he shows that using the same regression methodology people use to argue value spreads predict the value premium, you can “predict” the performance of equity anomalies using completely absurd variables — sunspot activity, the alignment of the planets, and global warming trends:

“Sunspots are an incredibly powerful predictor of momentum and post-earnings announcement drift.”

The point: time-series regressions used to “predict” factor premia are deeply flawed. With persistent regressors and overlapping observations, spurious correlation is almost guaranteed. This paper was inspired by a 1970s–80s literature that claimed interest rates and the yield curve could predict market performance — claims that failed out of sample. Novy-Marx saw the same methodology being recycled: “I saw people doing this again with other anomalies, and so I wrote this paper.”

Together, Dai’s 720-strategy analysis and Novy-Marx’s sunspot paper form a powerful one-two punch against factor timing. Valuation ratios are useful for identifying cross-sectional differences in expected returns (cheap vs. expensive stocks at any point in time), but not for time-series predictions (when to be in or out of factors).

7.5 Bessembinder: Why Broad Diversification Is Non-Negotiable¶

Hendrik Bessembinder (Arizona State University) published landmark research that reshapes how investors think about individual stock returns and the case for diversification. In an interview on Rational Reminder (Episode 346, February 2025), he discussed findings with sweeping implications.

The Extreme Skewness of Stock Returns

Using the full CRSP database (1926–present, ~26,000 stocks):

Roughly 57% of stocks destroyed shareholder wealth, another ~39% collectively earned just enough to offset that destruction, and 4% of stocks generated all of the net wealth creation above Treasury bills since 1926.

Bootstrap Simulations:

Even a randomly selected 100-stock portfolio underperforms the market more than half the time over a 90-year horizon. Skewness is diminished by diversification but never fully eliminated.

“A single stock strategy only beats the market 4% of the time. I bounced this question off some of my colleagues — their guesses were over 50%.”

Where Skewness Comes From:

1. Limited liability creates asymmetry: stocks can lose at most 100%, but gains are uncapped

2. Compounding mechanics amplify skewness: up-then-down sequences slightly lose money, while consecutive up-moves compound multiplicatively. Higher volatility makes these effects stronger.

The Skewness Tax Is Larger Than Fees

In a study of ~8,000 US equity mutual funds, Bessembinder found that to get 50% of funds to beat SPY, you’d need to add 12 basis points per month to every fund — that’s the pure skewness penalty. Average fees are only ~9 basis points/month. Even if active management were free, the majority of funds would still underperform because of the mathematics of skewness.

Implications for Factor Investors:

1. Concentrated factor portfolios face the same skewness problem as concentrated stock picks — missing the handful of top performers devastates returns

2. Broad factor tilts (holding thousands of tilted stocks, as DFA and Avantis do) hedge against this inevitability

3. The case for broad, systematic factor implementation is strengthened relative to concentrated stock-picking

4. Don’t focus solely on the mean — the median investor outcome is below the mean due to skewness

7.6 The Honest Verdict¶

The intellectually honest position — consistent with the full body of evidence — is probably this:

1. Chen is right that most published anomalies are worthless after costs in the modern era, and that economic theory adds less value than the profession claims

2. He is probably too pessimistic about well-implemented multi-factor strategies, which operate very differently from academic long-short replication

3. The expected premium is likely smaller than historical averages suggest — perhaps 1–2% annualized rather than 3–5%

4. The premium is uncertain enough that investors who cannot tolerate 10–15 years of potential underperformance should probably stick with total-market indexing

5. The method matters more than the story — how you implement (integrated, patient, low-cost) matters more than which theoretical narrative you find compelling

6. Constant adaptation may be necessary — the “buy and hold factor tilts forever” approach may need updating as markets evolve. This is uncomfortable for passive-minded investors.

What Chen gets right that everyone should internalize:

Key references:

• Dai, W., “Premiums for the Long Run” (Dimensional, 2022)

• Dai, W. & Medhat, M., “How Should Investors Set Return Expectations?” (Dimensional, 2023)

• Dai, W. & Novy-Marx, R., “Reversals and the Returns to Liquidity Provision” (Financial Analysts Journal, 2024)

• Dai, W., Medhat, M. & Novy-Marx, R., “On the Valuation of Performance Fees” (2022)

• Bessembinder, H., “Do Stocks Outperform Treasury Bills?” (Journal of Financial Economics, 2018)

• Bessembinder, H., “Which U.S. Stocks Generated the Highest Long-Term Returns?” (Financial Analysts Journal, 2023)

• Bessembinder, H., “Mutual Fund Performance at Long Horizons” (2024)

• DeMiguel, V. et al., “A Transaction-Cost Perspective on the Multitude of Firm Characteristics” (RFS, 2020)

• O’Reilly, G., Rational Reminder Episode 360: “The Components of Net Returns” (June 2025)

• Repetto, E., Rational Reminder Episode 315: “An Update from Avantis” (July 2024)

• Novy-Marx, R., “Predicting Anomaly Performance with Politics, the Weather, Global Warming, Sunspots, and the Stars” (2014)

• Asness, C., “The Long Run Is Lying to You” (AQR, 2020)

Section 8: Rolling-Window Analysis — Time-Varying Factor Exposures¶

Why Rolling Windows?¶

The regressions so far estimate constant factor loadings over the full sample. But in the Time Series Foundations notebook we saw that financial data can exhibit regime changes and structural breaks — the statistical properties of returns are not necessarily stable over 20+ years. Factor exposures can change over time due to:

• Changes in the index methodology or fund holdings

• Market regime shifts (e.g., value rotation vs. growth dominance)

• Varying sector composition within size/value groups

• Volatility clustering (Section 4 of the Time Series notebook) — high-volatility regimes may alter factor sensitivities

Rolling-window regressions fit the model over a moving window (e.g., 36 months) and let us see how betas evolve. This is a powerful diagnostic tool and is used extensively by practitioners. It connects directly to the stationarity question from Section 2 of the Time Series notebook: if factor loadings are non-stationary, a full-sample regression gives us an average loading that may not reflect the current regime.

Why plain OLS inside each window? In the 3-Factor tutorial we used Newey-West standard errors to account for autocorrelation over the full sample. Inside each 36-month rolling window, the primary goal is estimating the point estimates (betas), not conducting hypothesis tests. OLS point estimates are unbiased regardless of the error structure (Section 5 of the Statistical Foundations notebook), so Newey-West would change only the standard errors, not the estimated betas. Moreover, with only 36 observations the Newey-West long-run variance estimate becomes unreliable (few lags, few observations per lag). For these reasons, practitioners typically use plain OLS inside rolling windows and reserve HAC corrections for the full-sample inference.

Section 9: Summary and Further Reading¶

What We’ve Learned¶

This notebook covered the major extensions and debates surrounding factor investing:

The Factor Model Hierarchy¶

Level 1: CAPM (1 factor)
   R_i - R_f = α + β_MKT(R_m - R_f) + ε
   
Level 2: Fama-French 3-Factor (1993)
   + SMB (size) + HML (value)
   
Level 3: Carhart 4-Factor (1997)
   + UMD (momentum)
   
Level 4: Fama-French 5-Factor (2015)
   + RMW (profitability) + CMA (investment)
   [note: momentum intentionally excluded by Fama-French]
   
Level 5: FF5 + Momentum (6-Factor, practitioner standard)
   The "kitchen sink" model used in much applied research
   
Alternative: q-Factor Model (Hou, Xue, Zhang)
   MKT + ME (size) + ROE (profitability) + I/A (investment)
   Motivated by production-based theory

Looking Back & Looking Forward¶

This notebook is part of a five-notebook progression:

The key insight that runs through all notebooks: factor investing is not about chasing anomalies — it is about understanding the economic mechanisms that generate premia, measuring them with sound statistical tools, and implementing them in a way that survives transaction costs and behavioral pitfalls.

The next notebook puts this into practice by analysing the investment strategies of Dimensional and Avantis, running factor regressions on their European UCITS ETFs, and constructing a model portfolio.

Essential Reading¶

Foundational Papers:

• Fama & French, “Common Risk Factors in the Returns on Stocks and Bonds” (1993)

• Fama & French, “A Five-Factor Asset Pricing Model” (2015)

• Carhart, “On Persistence in Mutual Fund Performance” (1997)

• Novy-Marx, “The Other Side of Value: The Gross Profitability Premium” (2013)

• Black, Jensen & Scholes, “The Capital Asset Pricing Model: Some Empirical Tests” (1972)

• Frazzini & Pedersen, “Betting Against Beta” (2014)

Modern Research:

• Chen & Zimmermann, “Open Source Cross-Sectional Asset Pricing” (2022)

• Chen, Lopez-Lira & Zimmermann, “Does Peer-Reviewed Research Help Predict Stock Returns?” (2024)

• Chen & Velikov, “Zeroing in on the Expected Returns of Anomalies” (2023)

• Novy-Marx & Velikov, “A Taxonomy of Anomalies and Their Trading Costs” (2016)

• Novy-Marx & Velikov, “Betting Against Betting Against Beta” (2022)

• Novy-Marx, “Predicting Anomaly Performance with Politics, the Weather, Global Warming, Sunspots, and the Stars” (2014)

• Hou, Xue & Zhang, “Replicating Anomalies” (2020)

• Daniel & Moskowitz, “Momentum Crashes” (2016)

• Asness, Moskowitz & Pedersen, “Value and Momentum Everywhere” (2013)

• Cederburg & O’Doherty, “Does It Pay to Bet Against Beta?” (2016)

• Hong & Sraer, “Speculative Betas” (2016)

Practitioner Research:

• Dai & Novy-Marx, “Reversals and the Returns to Liquidity Provision” (Financial Analysts Journal, 2024)

• Dai, “Another Look at Timing the Equity Premiums” (Dimensional, 2023)

• Dai, Medhat & Novy-Marx, “On the Valuation of Performance Fees and Their Impact on Asset Managers’ Incentives” (2022)

• Bessembinder, H., “Do Stocks Outperform Treasury Bills?” (Journal of Financial Economics, 2018)

• Bessembinder, H., “Which U.S. Stocks Generated the Highest Long-Term Returns?” (Financial Analysts Journal, 2023)

• Bessembinder, H., “Mutual Fund Performance at Long Horizons” (2024)

Books:

• Andrew Ang, Asset Management: A Systematic Approach to Factor Investing (2014)

• Turan Bali, Robert Engle & Scott Murray, Empirical Asset Pricing: The Cross Section of Stock Returns (2016)

• Antti Ilmanen, Expected Returns (2011)

• Andrew Berkin & Larry Swedroe, Your Complete Guide to Factor-Based Investing (2016)

Exercises¶

Test your understanding of the advanced topics covered in this notebook. These exercises extend the hands-on skills from the 3-Factor Tutorial exercises — if any of the regression mechanics feel unfamiliar, revisit the earlier notebook first.

Exercise 1: 5-Factor Model on a New ETF (Hands-On)¶

Download return data for IWM (iShares Russell 2000 ETF — small-cap blend, no explicit value tilt) and run the full 5-factor regression.

1. How does IWM’s SMB loading compare to IWN’s? Is this what you expected?

2. Does IWM have a significant HML loading? Why or why not?

3. What happens to IWM’s alpha when you move from 3 factors to 5 factors?

# Starter code:
raw = yf.download('IWM', start='2000-01-01', end='2023-12-31', progress=False)
# ... compute monthly excess returns, then:
# res3 = run_ff_regression(None, ['Mkt_RF', 'SMB', 'HML'], iwm_df)
# res5 = run_ff_regression(None, ['Mkt_RF', 'SMB', 'HML', 'RMW', 'CMA'], iwm_df)

Exercise 2: Rolling Sharpe Ratios for Combined Strategies (Hands-On)¶

Using the multi-factor strategy returns computed in Section 7:

1. Compute the rolling 36-month Sharpe ratio for (a) HML alone, (b) Value + Profitability, and (c) Value + Profitability + Momentum.

2. Plot all three on the same chart. During which periods does the combined strategy most clearly outperform single factors?

3. Is there any period where all three strategies have a negative rolling Sharpe ratio simultaneously?

# Starter code:
# rolling_sharpe = ret_series.rolling(36).mean() / ret_series.rolling(36).std() * np.sqrt(12)

Exercise 3: Momentum Crash Anatomy (Discussion + Hands-On)¶

1. Identify the worst 3-month period for the momentum factor (UMD) in our sample. What calendar months were they?

2. What was the market (Mkt-RF) doing during those months? Was this consistent with the Daniel & Moskowitz (2016) narrative of momentum crashing during market recoveries?

3. If you had held an equal-weight combination of UMD and HML, how much would the drawdown have been reduced?

# Starter code:
# rolling_3m_mom = factors['UMD'].rolling(3).sum()
# worst_end = rolling_3m_mom.idxmin()
# print(f"Worst 3-month period ending: {worst_end}")

Exercise 4: The Beta Anomaly — Subsample Robustness (Discussion + Hands-On)¶

The beta anomaly (flat or negative SML) was documented in data going back to the 1930s. But does it hold in every subperiod? This connects to the stationarity discussion in the Time Series Foundations notebook — if the anomaly is non-stationary, our full-sample estimates may be misleading.

1. Split the beta-sorted decile portfolios into two halves: 2000–2011 and 2012–2023. Plot the empirical SML for each subperiod. Is the anomaly consistent, or does it appear mainly in one era?

2. Regress the simple BAB factor (D1 − D10) on FF5 + Momentum separately for each subperiod. Does the alpha change sign or significance?

3. Discussion: Frazzini and Pedersen argue leverage constraints drive the BAB premium. Did anything change about leverage availability (e.g., the rise of leveraged ETFs, zero-commission brokers) between the two subperiods that might affect the premium?

# Starter code:
# beta_first = beta_ports.loc['2000':'2011']
# beta_second = beta_ports.loc['2012':'2023']
# For each half, recompute avg excess returns and realized betas, then plot.