Perspectives · September 2026

A prospective client came to us this year with a portfolio he had built himself over about two decades. Twenty-seven individual companies. He is a businessman, and he had picked them the way a businessman would: profitable firms, run by people who seemed to know what they were doing, bought at prices that seemed fair. Two of them had been extraordinary. One was up roughly fifteen times from his cost. Another more than seven.

He wanted to know whether the approach was working. So did we. We ran his portfolio through the same factor framework that academics have used for thirty years to answer exactly that question. The result was not what I expected, and it has changed how I plan to have this conversation with the next person who asks.

What the test actually does

The idea behind factor analysis is straightforward. Most of what any diversified stock portfolio does is explained by a handful of common characteristics. How much market exposure it has. Whether it leans toward small companies or large ones. Whether it leans toward cheap stocks or expensive ones. Whether the companies are profitable. Whether they invest conservatively or aggressively.

Run a portfolio's monthly returns against those five factors and you get a number for each one, called a loading. A loading tells you how much of that characteristic the portfolio actually has. Whatever return is left unexplained is called alpha.

This matters because the characteristics are not free money. They are compensated risks. Decades of evidence say that small companies, cheap companies, and profitable companies have delivered higher returns over long periods, and that investors are paid for holding them. If an investor believes he is buying profitable companies at fair prices, the regression will say whether he actually is.

What we found

Here is the five-factor result for the twenty-seven stocks, using nine years of monthly returns.

Fama-French five-factor regression, monthly, nine years ending June 2026. Significance assessed at the five percent level.

Read the bottom four rows. Not one of them is distinguishable from zero. The portfolio has no size exposure, no value exposure, no profitability exposure, and no investment exposure. What it has is market risk, slightly more than one for one.

The profitability line is the one that stopped me. Profitability is precisely what this investor believed he was selecting for. He described it clearly and he described it correctly — the research supports the idea. His portfolio has a loading of negative two hundredths, with a t-statistic of negative thirty-five hundredths. Statistically, it isn't there at all.

Why it isn't there

The reason turns out to be arithmetic rather than judgment, and once you see it you cannot unsee it.

He owned a large integrated oil producer, which loads +1.13 on value. He also owned a mega-cap online retailer, which loads −0.45. He owned a regional bank at +1.16 and a mega-cap software firm at −0.47. He owned midstream energy partnerships in the +0.70 to +1.00 range and a consumer technology giant at −0.49.

Each of those was a defensible individual decision. Held together, in roughly equal size, they offset. The value stocks and the growth stocks cancel, and the portfolio nets to nothing.

The same thing happens on profitability. Three consumer and industrial names loading between +0.65 and +0.82 on one side. A high-growth software company at −1.57, a pharmaceutical firm at −0.48 and a heavy equipment manufacturer at −0.43 on the other. Net: zero.

This is not a story about bad stock picking. It is a story about what happens when decisions are made one at a time, over twenty years, without anyone ever standing back and measuring the whole.

The alpha, and why I do not trust it

The regression also produced an annual alpha of 8.22%, with a t-statistic of 4.42. That is a large number and it is statistically significant. An investor would be entitled to point at it.

I would be careful, for three reasons.

The weights are set by the outcome. The regression uses today's allocation. One position is nearly twenty percent of that portfolio and another sixteen — not because he sized them that way, but because they went up fifteen times and seven times respectively. Running today's weights backward through the nine years that made them large is a backtest in which the biggest positions are biggest precisely because they won. The portfolio being measured is not the portfolio he held in 2017.

The model is missing the factor most likely to explain it. The Fama-French five-factor model contains no momentum term. A book of mega-cap winners held through 2017 to 2026 is a momentum portfolio whether or not anyone intended it. I would expect a meaningful share of that alpha to move into a momentum loading if the factor were included.

It does not appear at the individual security level. Of twenty-six stocks with sufficient history, exactly two showed statistically significant alpha, and together they were under six percent of the portfolio. The largest holding, at a fifth of the book, had a t-statistic of 1.36 — not significant. Five positions carried negative alpha. If there were a repeatable skill at work, it should show up in more than two names out of twenty-six.

None of this means his returns weren't real. They were. It means the evidence cannot tell us whether they were skill, and that the mechanism that produced them — two positions becoming a third of the portfolio — runs in both directions.

The funds told the opposite story, and it was just as unhelpful

The same household held twenty-seven mutual funds and ETFs alongside the stocks. We ran those too.

Four of the household's largest actively managed equity funds. Same model, same period.

R-squareds between 0.94 and 0.98, and not one alpha distinguishable from zero. Five factors explain virtually everything these funds did. Three other funds in the household carried statistically significant negative alpha.

Meanwhile about forty percent of the fund sleeve sat in gold, uranium and platinum, whose R-squareds were 0.09 and 0.26. Three quarters or more of what those positions did was unexplained by anything the model measures.

So the household held two piles. A stock book that looked active and was, in aggregate, the index. And a fund book where the actively managed portion was the index priced as active management, and the rest was commodity exposure nobody had sized deliberately.

What I take from it

One. Intent is not exposure. The investor's stated approach was sound and matched the research. His portfolio did not contain it. The only way to know the difference is to measure.

Two. Offsetting bets are the quiet failure mode. Concentration risk is visible; anyone can see that two stocks are a third of a portfolio. Cancelled factor exposure is invisible. You cannot find it by looking at a statement, and it costs you the premium you thought you were buying while leaving all the single-name risk in place.

Three. Alpha measured on today's weights is not a track record. It is a description of which positions won, dressed as a finding. Ask what the portfolio looked like at the start of the period before you believe it.

Four. This is a structural problem and it has a structural fix. If an investor wants exposure to profitability, value and size, those characteristics can be bought deliberately, in known amounts, across thousands of companies — and measured again a year later to confirm they are still there. That is a different thing from hoping that twenty-seven individual judgments happen to average out the right way.

If you have not read it, the companion to this piece is The Planning Problem, which takes up the other half of the question: not what a portfolio is exposed to, but over what horizon that exposure should be judged.

A note on method

Loadings and alphas come from ordinary least squares regressions of monthly excess returns on the Fama-French factors, over nine years ending June 2026. Portfolio-level regressions use current position weights. Individual security regressions use each security's full available history within the window. Statistical significance is assessed at the five percent level. Holdings are described generically and figures are rounded. The analysis is drawn from a single household's portfolio and is presented for illustration; different portfolios, periods and models will produce different results.