CAPM: What It Explains and What It Doesn’t

Analyst Desk

The Investor Problem

Every investor prices risk, whether or not they know it. Demand a higher return from a volatile mining stock than from a utility stock, and you have asserted a theory of how risk and reward relate. The question is whether your theory is coherent. For over sixty years, the starting point for every coherent answer has been the Capital Asset Pricing Model.

The CAPM, developed in the early 1960s by William F. Sharpe (building on Markowitz’s portfolio theory, and independently by Lintner, Mossin, and Treynor), earned Sharpe a Nobel Prize and became the most widely taught and most empirically battered model in finance. Practitioners use it daily to estimate costs of capital; academics have spent four decades documenting its failures. Both camps are behaving rationally, and understanding how that can be, what the model genuinely explains, and where it breaks, is one of the most useful pieces of intellectual equipment an analyst can own.

Only Undiversifiable Risk Gets Paid

The CAPM’s central insight survives every empirical assault on its details, and it is this: the market does not pay you for risks you could have diversified away.

Consider a single stock. Its price swings for two kinds of reasons: things specific to the company (a factory fire, a product win, a bad manager) and things that move the entire market (recessions, interest rates, panics). The specific risks, as Month 2 established, vanish in a diversified portfolio. Your factory fire and my product win cancel. Since any investor can hold a diversified portfolio, no rational buyer demands extra return for bearing specific risk; the only risk that survives diversification, and therefore the only risk that commands a premium, is the co-movement with the market itself.

The measure of that co-movement is beta: how much a stock tends to move when the market moves. A beta of 1 means the stock swings with the market; 1.5 means it amplifies market moves by half again; 0.5 means it dampens them. The model then makes a startlingly precise claim:

E(Rᵢ) = Rf + βᵢ × (E(Rm) − Rf)

An asset’s expected return equals the risk-free rate plus its beta times the market risk premium. Nothing else matters, not the company’s total volatility, not its size, not its story. Expected return is a straight line in beta: the Security Market Line. Double an asset’s beta and you double its risk premium. A stock that gyrates wildly for purely company-specific reasons deserves, in this framework, no more return than a government bond, provided its beta is zero.

This is a genuinely deep result. It says risk is not a property of an asset in isolation, but of its relationship to everything else. The right question is never “how risky is this?” but “what does this add to a portfolio?”

What the Model Explains Well

Before the criticisms, let’s give credit where due. The CAPM’s qualitative content is broadly right and permanently useful.

It explains why diversification is unrewarded-risk removal. The distinction between systematic and idiosyncratic risk is the model’s skeleton, and it survives intact in every successor model. An investor who understands only this has understood the most important thing.

It disciplines the cost of capital. Ask what return a business must earn to justify its risk, and the CAPM supplies a defensible, auditable answer: the risk-free rate plus a beta-scaled premium. Corporate finance runs on this arithmetic and most large firms actually estimate hurdle rates this way, not because the number is precisely right, but because it is transparent: every input can be argued about openly, which is more than can be said for intuition.

It created honest performance measurement. Once expected return depends on beta, a manager’s raw return can be decomposed: how much came from taking market exposure (cheaply available to anyone), and how much was genuine excess (alpha). Jensen’s alpha, the Sharpe ratio, and the entire apparatus of risk-adjusted evaluation (a later month’s subject) descend from this. Before CAPM, a manager who bought high-beta stocks in a bull market looked like a genius; after it, he looked like a man renting leverage and billing for skill.

What It Doesn’t Explain

The model’s precise quantitative claim; the assumption that expected returns line up on beta, and only beta has failed nearly every careful test, and the failures are instructive rather than embarrassing.

The line is too flat. Across decades of data, low-beta stocks have earned more than the model predicts and high-beta stocks less. The measured Security Market Line is far flatter than theory demands. This “low-beta anomaly” is among the most robust patterns in finance, plausibly rooted in leverage constraints (investors who cannot borrow instead overpay for high-beta stocks as embedded leverage) and in the lottery-ticket appeal of volatile names.

Beta is not the only priced variable. Fama and French’s landmark 1992 tests found that once size and value characteristics are counted, beta’s standalone explanatory power over the cross-section of returns nearly vanishes. Small stocks and cheap stocks earned more than their betas justified. This finding launched the multi-factor era (and a later month of this blog). Momentum deepened the wound.

Its assumptions are heroic. A single period; identical investor expectations; frictionless borrowing and lending at one risk-free rate; a “market portfolio” containing all wealth; including land, private businesses, and human capital, which, as Richard Roll’s famous critique observed, makes the model strictly untestable, since every empirical test uses a stock index as a stand-in for a market portfolio no one can observe. An investor in Kampala whose true wealth is mostly land, a business, and future earnings holds a “market portfolio” no index fund resembles and therefore her priced risks differ accordingly.

And beta itself is unstable. Estimated from noisy history, drifting across regimes, sensitive to the index and window chosen. Next week’s Quant Lab estimates betas by hand precisely to make this instability visible.

Limitations and Real-World Complexity

How, then, should a practitioner hold the model? The professional consensus has settled into a mature two-mindedness. As a predictive pricing formula, the CAPM is unreliable. Therefore, no serious quant prices the cross-section of stocks on beta alone, and using it for that is malpractice. As a conceptual framework and communication standard, it remains indispensable: cost-of-capital estimation still starts there (often with factor or size adjustments bolted on); performance attribution still speaks its language; and its core distinction; priced systematic risk versus unpriced diversifiable risk, is the foundation on which every successor model, including the multi-factor ones that dethroned it, still stands. The factor models did not refute the CAPM’s logic; they multiplied it, replacing one systematic risk with several.

There is also a philosophical caution the model teaches by negative example: a theory can be axiomatically beautiful, Nobel-worthy, universally taught and empirically wrong in its precise claims. Finance is not physics. Its models are lenses, not laws.

The Long-Term Lesson

The CAPM’s enduring gift to the thoughtful investor is not its equation but its questions, which remain the right ones to ask of any holding: How much of this asset’s risk would survive diversification? Is the expected return compensation for that surviving risk, or am I hoping to be paid for risk the market gives away free? And if a manager claims outperformance, how much is market exposure in costume?

Ask those three questions habitually and you are using the CAPM as its authors intended (as an instrument for thinking clearly about the price of risk). Expect its number to be precise, and you will join a long line of the disappointed. The model explains the structure of risk and reward about as well as anything in social science explains anything. What it cannot do, what no model has yet done, is dispense with judgment about the particulars. The equation starts the analysis. It has never once finished it.