Alpha vs Factor Exposure

You can now run a factor regression and read out a fund’s loadings and its intercept. This lesson is about what that intercept means — and why it is the single most misunderstood number in all of investing. A fund returns 14% in a year the risk-free rate was 3%; it beat the S&P; the manager is on magazine covers. Did they earn that with skill, or did they just tilt their portfolio toward small, cheap, recently-winning stocks and collect the premia those tilts have always paid? Factor models exist to answer exactly that question, and the answer is usually deflating: most of what looks like genius is rentable beta, and the sliver that’s left — the true alpha — is frequently negative once you benchmark honestly.

The return decomposition — skill plus tilt

Analogy. Picture a chef’s restaurant bill. The headline total looks impressive, but an itemized receipt shows most of it is ingredients you could have bought yourself at the market — flour, eggs, butter — priced at the going rate. Only the last line, “chef’s labor,” is what you’re actually paying the chef for. A factor model is that itemized receipt for a return: it lists how much of the total came from generic, rentable risk exposures, and leaves the true skill as the final line.

Definition. A factor model says an asset’s expected excess return (return above the risk-free rate $R_f$) is the sum of its exposures to systematic factors, each priced by that factor’s long-run premium, plus a residual skill term:

$E[R_i] - R_f = \sum_{k} \beta_{i,k}\,\lambda_k + \alpha_i.$

Here $\beta_{i,k}$ is the asset’s loading on factor $k$ (how much it moves with that factor), $\lambda_k$ is the factor’s premium (the long-run reward per unit of loading), the product $\beta_{i,k}\lambda_k$ is that factor’s contribution to the return, and $\alpha_i$ is the intercept — the part no factor explains. That residual is the only piece that can plausibly be called manager skill.

The crucial dynamic: every factor you add to the model is another way to explain away apparent alpha. A return is a fixed total; the more of it your factors absorb, the less is left for the intercept. Alpha isn’t a property of the fund alone — it’s a property of the fund relative to the benchmark model you chose. Change the model and the same fund’s alpha changes.

Alpha is often hidden beta

Analogy. A magician’s “levitation” looks supernatural until someone hands you better lighting and you spot the wire. Nothing about the trick changed — your instrument got sharper, and the magic resolved into mechanism. Apparent alpha is the levitation; adding the right factor is the better lighting; the wire is a loading you’d been missing.

Definition. Apparent alpha under a sparse model (few factors, e.g. CAPM’s single market factor) routinely shrinks — often to zero or below — under a richer model (many factors). The mechanism is omitted-variable bias. Suppose a fund loads positively on a factor $F$ that has a positive premium $\lambda_F$, but your model leaves $F$ out. The regression still has to fit the data, so it dumps that missing contribution $\beta_F \lambda_F$ into the only free parameter it has left: the intercept. You read a fat alpha. Add $F$ as a regressor and that chunk of return migrates from the intercept into the $\beta_F$ loading — alpha collapses by exactly $\beta_F \lambda_F$. Nothing about the fund changed; the benchmark just got smarter and stopped crediting the manager for a tilt anyone could buy.

This is why factor research has a recurring obituary structure: a celebrated “alpha” is published, a new factor is identified, and the alpha is revealed to have been that factor’s beta in disguise.

A famous case: much of Warren Buffett’s legendary “alpha” turns out, under a rich factor model, to be exposure to quality (profitable, stable, well-managed firms), low volatility, and cheap leverage applied to those tilts — Frazzini, Kabiller and Pedersen (2018) show Berkshire’s record is largely explained, not mystical, once you regress it on the right factors. It doesn’t make Buffett less remarkable — he found and levered durable factors decades before they were named — but it relocates the achievement from “unexplainable genius” to “early, disciplined factor exposure.” Earlier, Carhart (1997) did the same demolition to mutual funds: most funds with glowing CAPM alphas were simply riding the momentum factor, and once UMD was added their alphas vanished.

The Information Ratio — skill per unit of active risk

Analogy. Two students both raise the class average by five points, but one did it with steady, reliable work and the other with wild, unpredictable swings that happened to net positive. You trust the steady one more. The information ratio rewards consistent outperformance: a given alpha earned with small, stable deviations from the benchmark is worth far more than the same alpha earned with violent ones.

Definition. The information ratio (IR) divides a fund’s alpha by its tracking error — the standard deviation of its residual (active) return relative to the benchmark:

$\text{IR} = \frac{\alpha}{\sigma(\varepsilon)} = \frac{\text{alpha}}{\text{tracking error}}.$

It measures skill per unit of active risk: how much excess return the manager extracts for each unit of how far they’re willing to stray from the benchmark. Contrast it with the Sharpe ratio, which divides excess return by total risk. Sharpe asks “is the whole portfolio efficient?”; IR asks the sharper question “is the manager’s deviation from the benchmark paying off?” A closet-indexer with tiny tracking error needs only a tiny alpha to post a great IR; a cowboy with huge active bets needs a huge alpha to look good.

Rough benchmarks. A sustained IR above about 0.5 is genuinely good; an IR above 1.0, maintained over many years, is rare and elite — the territory of the best active managers alive. And the sign matters as much as the size: a negative IR means the manager’s active bets are destroying value relative to just holding the benchmark.

A full worked attribution

Now the payoff: take the magazine-cover fund and itemize its receipt. The fund returned 14% in a year the risk-free rate was 3%, so its excess return is 11% (14% − 3%). We regress it on the four Carhart factors and read off the loadings; we multiply each loading by that factor’s long-run premium to get its contribution.

The factors alone explain 11.65% of excess return — more than the fund actually delivered. So the true alpha is the leftover:

$\alpha = 11\% - 11.65\% = -0.65\%.$

The fund underperformed its factor-replicating benchmark. You could have cloned its strategy with cheap factor ETFs — a dab of market, small-cap, value, and momentum exposure in those proportions — and earned 11.65%, beating this manager by 0.65 percentage points while paying a fraction of the fees. The “outperformance” versus the S&P was entirely the size, value, and momentum tilts; the manager’s own contribution was slightly negative.

Push it one step further with the information ratio. Say the fund’s residual tracking error is 4%. Then:

$\text{IR} = \frac{\alpha}{\sigma(\varepsilon)} = \frac{-0.65\%}{4\%} = -0.16.$

A negative IR confirms it: the manager’s active bets, net of factor exposure, subtracted value. This is the whole course in one number — a fund that beat the index, beat the risk-free rate, and looked like a winner is, under an honest benchmark, a slightly skill-negative factor portfolio in an expensive wrapper.

Why “beating the S&P” usually isn’t alpha

Analogy. Bragging that you outran a tortoise tells me little about your speed — it tells me you raced a tortoise. The S&P 500 is a cap-weighted index: a specific bet on large-cap stocks, weighted by size. Beating it while tilted toward small, cheap, or recently-winning stocks isn’t beating the market of risk — it’s racing a benchmark that happens to be slow on the very dimensions you tilted toward. You collected premia; you didn’t generate skill.

Definition. True alpha is what’s left after benchmarking to your actual factor exposures, not to a convenient cap-weighted index. The right comparison for a small-value-momentum fund is a portfolio with the same small, value, and momentum loadings — bought cheaply through factor ETFs. Only the return above that is skill. “Beat the S&P” is a marketing claim; “positive alpha against a matched factor benchmark with a positive information ratio” is a skill claim, and they are wildly different bars.

Trade-off — the benchmark is a judgment call. A richer model is a fairer skill test, but you can over-control. If a manager’s genuine edge is an early, well-timed bet on some factor, and you include that factor in the benchmark, you’ll subtract away the very thing they should be credited for — declaring a real edge to be zero alpha. (Imagine benchmarking Buffett against a quality factor that didn’t exist when he started exploiting it.) So choosing which factors belong in the benchmark is itself a judgment: too few and you credit rentable beta as skill; too many and you erase a manager’s real, hard-won exposure. There is no model-free answer — only a defensible choice you must be able to argue for.

Putting it together

A return splits cleanly into tilt and skill: $E[R_i] - R_f = \sum_k \beta_{i,k}\lambda_k + \alpha_i$, where each factor contributes its loading times its premium and alpha is the residual no factor explains. Because the total is fixed, every added factor can only shrink apparent alpha — and much of what passes for skill is hidden beta, an omitted factor’s contribution that the regression dumped into the intercept (Buffett’s quality-and-leverage record, Carhart’s momentum-riding funds). The information ratio, alpha over tracking error, measures skill per unit of active risk, with sustained values above 0.5 good and above 1.0 elite. The worked attribution made it concrete: a 14% fund beating the S&P had its entire 11% excess explained by factor tilts worth 11.65%, leaving a true alpha of −0.65% and an information ratio of −0.16 — a cheap ETF clone would have won. The lesson: beating a cap-weighted index isn’t alpha; alpha is what survives benchmarking to your actual exposures — and choosing that benchmark, neither too sparse nor too rich, is the analyst’s real judgment call.

That closes the loop from estimation to verdict: you can now decompose any return into rentable factor exposure plus a residual, judge that residual with the information ratio, and refuse to be impressed by “beat the S&P.” The final lesson, smart beta and the factor zoo, turns the same skepticism on the factors themselves — because if alpha is mostly hidden beta, the next question is how many of those “factors” are even real.