A factor is a rule; whether the rule pays is a separate question

A factor is a characteristic a rule can sort on

A factor is a measurable characteristic shared by a group of securities: how small the company is, how cheaply it is priced against its accounts, how it has moved recently, how profitable it is, how much it bounces around. A rule can sort a universe on any of them without anyone forming a view about a single company.

Take a universe of 500 listed companies. Rank them by market value, keep the smallest hundred, hold those for a year, then rank again. That is a complete investment process. It has no opinion about management quality, no meeting with a chief executive, no judgement of any kind in it — and anyone handed the same data and the same instruction produces the same hundred names.

Swap the ranking variable and you have a different process with the same property. Rank by price against book value and keep the cheapest hundred. Rank by the past year's price change and keep the strongest hundred. Rank by return on capital and keep the most profitable hundred. In each case the characteristic being sorted on is doing all the work, and the characteristic has a name: a factor.

That is the whole idea, and it is worth noticing how modest it is. A factor is a defined ranking variable, nothing more. It carries no promise about what the sorted portfolio will earn, and most of this article is about the distance between those two things.

The five that get named, and what each sort leaves out

Five characteristics come up often enough to have settled names. The table states what each rule actually measures and, in the third column, what the measurement cannot see — which is where most of the trouble lives.

Now the mistake that follows directly from that table, and it is the commonest one in this whole subject: reading a factor label on a product as a factor definition. Value is not one variable. Book-to-price, earnings yield, cash-flow yield and sales-to-price each rank the same 500 companies differently, and a fund that says “value” has chosen one of them. The label does not say which.

The arithmetic of that is blunt. Two rules each keep the cheapest hundred of the same 500. Suppose the two lists agree on sixty names — a figure chosen to make the point land, not a measurement of anything. Then forty of each hundred are names the other rule looked at and rejected, so 40% of each portfolio is a holding the other definition would not own. Two funds, one label, substantially different books.

The same applies to every row. Momentum computed over twelve months and momentum computed over three months are different rules that will disagree at exactly the moments it matters. The distinction between the value and growth camps is set out in value versus growth, and the size buckets themselves in market capitalisation — neither is re-argued here.

Reproducibility is the part that is not in dispute

Here is the claim this article is willing to make flatly, because it follows from the structure rather than from any dataset: a rule can be written down, handed to a stranger and re-run to the same answer. A manager's judgement cannot. Everything else about factor investing is arguable; that is not.

Four consequences follow, and together they are the honest case for the approach.

• You can know what the portfolio will hold before it holds it, because the rule is the holding.
• You can recompute its record over any period you choose rather than the period someone selected for a marketing sheet — how much work the choice of window does is set out in rolling returns.
• You can tell whether a disappointing year was the rule doing what it said or someone quietly departing from it.
• Running a screen consumes less of what a research desk consumes, which is analyst time. Whether any of that saving reaches the holder is a question about how a product is priced rather than about the rule, and the expense ratio is where it would show up.

There is a fifth that is easy to undersell. A rule does not lose its nerve. It buys the cheap list in the month the cheap list is frightening, which is the month a person finds reasons not to — the mechanism behind that reluctance is crowd behaviour and its pull. Whether the buying is rewarded is unknown; that it happens on schedule is a property of the rule.

Now the cost, because nothing here is free. What reproducibility buys is the removal of judgement, and judgement was sometimes right. A value rule that ranks on book value will buy a company whose book is a plant that will never restart, and it will keep buying while the accounts still say the plant exists. A human analyst might notice. The rule's entire virtue is that it will not notice, and you cannot keep the discipline while carving out exceptions for the cases you dislike — the exceptions are the judgement, returning through the back door.

Three explanations for a premium, and they are not the same claim

Suppose a sort has, over some studied period, delivered more than the market it was drawn from. Three explanations are on offer, and they are usually presented as alternative flavours of one idea. They are not. They are different claims with different consequences.

• Compensation for a risk. The sorted portfolio holds something genuinely unpleasant — it falls harder in bad times, or falls when your job is also at risk — and the extra return is the wage for holding it. On this reading the return is not a free finding; it is a price.
• A persistent behavioural error. Enough investors systematically misprice the characteristic — extrapolating recent growth too far, avoiding names that have been embarrassing — and the rule collects the difference. On this reading the return exists because of a mistake other people keep making.
• An artefact of the sample. Nothing was there. The pattern is a property of the particular period and universe examined, and it is in the data because someone looked hard enough to find it.

Here is why the distinction is not academic. The three make different predictions about what happens once the finding is published, and that is the only test most readers will ever get to observe.

If the premium is payment for discomfort, publication changes nothing. Knowing that a holding hurts in bad times does not make it stop hurting, and nobody arbitrages away a wage for enduring something. If it is a behavioural error, publication may erode it — but only where capital can actually reach the trade, which excludes names too small or too illiquid to buy in size. If it is an artefact, there was never anything to erode, and the out-of-sample number reverts to what it always was.

There is a trap on the risk side that deserves naming, because it is where the debate usually goes soft. “It is compensation for risk” can be asserted after any outcome: the sort did well, so the risk did not show up; the sort did badly, so the risk did. A risk story earns its keep only when it says in advance which states of the world the holding is supposed to hurt in. Anything less is a label applied afterwards. And a characteristic that becomes a standard explanatory input stops being unexplained return at all — that mechanism, and what it does to a manager's headline figure, is alpha and beta.

What a record can settle, and what it cannot

Two kinds of statement get made about factors and they need different evidence. That a sort has a rationale — a story about why the characteristic should be priced — is a mechanism claim, and you check it by reading the argument. That the sort earned something is an empirical claim, and it needs a market, a period, a construction method and a cost assumption attached before it means anything. Historical figures describe what happened, not what will happen; past performance does not indicate future results.

The reason to be strict about that is a piece of arithmetic anyone can redo. A test of this kind asks whether a pattern is stronger than chance would ordinarily throw up, and the conventional bar sits where chance alone clears it one time in twenty — the 5% threshold. Now run 100 candidate characteristics against a price history that contains no real signal whatever. Roughly 100 × 0.05 of them, which is five, clear the bar anyway.

Publish those five. Every one arrives with a plausible story attached, because a plausible story can be constructed for any characteristic once you know it worked, and the ninety-five that failed are not in the paper. The survivors of an unseen search look identical to discoveries.

A second gap sits between the published number and anything a fund can hand you. A factor premium is usually measured as the spread between the top-ranked group and the bottom-ranked group — buy one, sell the other. A long-only fund cannot sell the bottom group; the least it can do is not own it, because a weight cannot go below zero. Take chosen figures: the top group ahead of the index by 2 percentage points and the bottom group behind it by 3. The long-short spread is the 5 points between them; the long-only version has access to 2, before any cost. Those inputs are picked to make the subtraction visible and describe no actual market.

Then there is the pattern the literature keeps reporting: documented premia have often looked smaller in the data arriving after the paper than in the data the paper used. Three readings compete — capital arrived and competed the return away, the finding was partly mined out of the sample to begin with, or the later window was an unlucky draw from the same distribution.

Treat that pattern as reported rather than as established, because this article gives you nothing to check it against. It names no magnitude for any factor in any market, and that is deliberate: a decay figure means nothing without the study it came from, the market it was measured in and the years it covers, and a US sample and an Indian one are not interchangeable. Those things belong in a citation, not in a sentence like this one.

The assumptions under the inference, not just under the rule

A model is a set of assumptions, and the assumptions are the model. The ones worth naming here are not the assumptions inside the rule — those belong to the fund and are set out in smart beta. They are the assumptions under the inference: the step from “this sort did well in this sample” to “this characteristic earns something”. Five of them, each known to be imperfect rather than merely unproven.

• A clean single-characteristic sort. There is no such thing. A portfolio of the smallest hundred companies is also, mechanically, the least liquid hundred, the least researched hundred and often the most leveraged. Attributing the outcome to size assumes every other exposure that travelled with it was noise, and the sort offers no way to check that.
• A survivor-free universe. A rule backtested on the companies listed today has quietly excluded everyone who delisted, merged or was suspended along the way — and those are disproportionately the failures. The sort never had to own them. A real investor running the rule in 2010 would have.
• An association worth acting on. The sort found that a characteristic and a return moved together in a sample. Using it forward promotes that correlation to a standing order, which assumes whatever produced the association is still operating. Nothing in the ranking tells you what that was, so nothing tells you whether it stopped.
• A single period. The framework asks what the portfolio earned between rebalance dates and treats the interval as one decision. A person holds it through every day in between, and the days in between are where holdings get abandoned. A rule that is correct and abandoned pays nothing.
• Mutually distinct sorts. Holding five rules assumes five distinct portfolios. Two rules ranking the same universe have no obligation to disagree — a quality screen and a low-volatility screen can converge on the same names — so the spread you thought you bought is a correlation assumption like any other.

Behind all five sits the statistical one, and it is the load-bearing assumption of the whole apparatus: that returns are draws from a stable distribution, independent enough that an average computed from them carries a meaningful error bar. Equity returns are not independent in that way. Their worst days arrive in clusters, so twenty years of monthly observations — 240 of them — hold rather fewer genuinely separate pieces of evidence than 240.

The error bar is nonetheless computed as though all 240 counted, which fixes the direction the failure runs in: too narrow, never too wide. Every backtest looks more settled than the data underneath it can support, and adding years narrows the true error bar more slowly than the count of months suggests.

None of this makes the approach unusable. What it makes it is conditional, and the conditions are checkable. The reason to name them rather than teach the framework clean is that a rule you can restate exactly feels like knowledge in a way a manager's hunch does not, and precision is not accuracy. A sort that is wrong is wrong reproducibly, on schedule, to five decimal places.

Four choices that decide what a factor label means

Everything above concerns the idea. A product is a separate object, and four construction choices sit between them. None appears on the front of the fund, and each one changes what the label is describing.

The first is the ranking variable, already covered: which measure of cheap, which definition of quality. The second is the measurement window. Momentum over twelve months and momentum over three months are different characteristics wearing one name, and the rebalance schedule decides how much of whichever one it is the portfolio actually holds at any moment — a rule reconstituted once a year spends much of that year holding names that stopped qualifying.

The third is weighting. Rank a universe, then weight the survivors by market value, and the result is dominated by the largest of them, which can leave very little of the characteristic in the portfolio. Equal weighting keeps more of it. The fourth is the cap structure — single-stock and single-sector limits — which converts a pure sort into a constrained one without changing the name on the front.

All four live in the index methodology document rather than the fund name, and asking for that document is the whole remedy. What it costs to run the resulting rule, and why a cap-weighted book is the only one that rebalances itself, belong to smart beta; the vehicle itself to index funds and exchange-traded funds; and where a scheme states what it does to reading a factsheet. One structural point is worth carrying anyway: the same rule run in a personal account realises a gain or a loss at every rebalance, where a scheme's internal trading hands you nothing until you redeem. The two arrangements settle at different moments.

There is a final cost that no document discloses, because it is temperamental rather than financial. A rule that departs from a cap-weighted index must sometimes sit behind it, and nothing says for how long. The depth and duration of those stretches are what drawdown measures exist to describe, and they decide whether the rule was still being followed when it finally mattered.

Testing a rule rather than arguing about it

The mechanism claims in this article are settled by reading. The empirical ones are not settled by reading at all, and the honest way to hold a view about a sort is to run it and look at what came out.

FNOTrader's Stocks app runs rank-and-hold rules over a universe of roughly 2,390 stocks and 17 NSE sector and size indices — rank on a chosen characteristic, hold a chosen number of names, rebalance on a chosen schedule — and reports the outcome in rupees rather than as an index number. Running the same rule over two different periods, and two different definitions of the same characteristic over one period, is the direct way to see how much of a result was the idea and how much was the choice. The Mutual Funds app covers the other side on around 34 million rows of net asset value — the per-unit price of a scheme — with rolling-return distributions across every start date available.

FNOTrader is not a SEBI-registered investment adviser or research analyst, and nothing here is a recommendation to buy, hold or avoid any scheme, index or security. No factor in this article is described as better than another, and none is claimed to earn anything. What the article offers is the list of questions a factor claim has to answer before it counts as evidence.

Common questions

Continue reading

More in Advanced Investing · App: Stocks · Definitions: glossary · Free tools: calculators · All: every article