What Happens When One Valuation Method Goes Wrong

Eight valuation methods run on a stock trading at 100.00. Seven of them land between 105 and 121. The eighth returns 480.

Nothing about that is exotic. A book value per share that arrived doubled, a share count read from the wrong column, a negative denominator that flipped a ratio — any of these produces one method that is confidently, spectacularly wrong while the other seven are fine. The question a blended fair value has to answer is not whether that ever happens. It is what the published number does when it does.

A plain average of those eight is 159.25. The seven good methods average 113.43. One bad input has moved the answer by 40.4%.

The blended midpoint our engine actually publishes for that same set moves from 107.45 to 110.38 — 2.7%. This is how, and what it costs.

Four defences, in the order they fire

The blend does not have one outlier rule. It has four mechanisms that fire in sequence, and which one catches a given bad method depends on how bad it is, what sector the company is in, and which other methods happen to have run.

1. The hard gate

Before anything statistical happens, every method fair value is compared to the current price and discarded if it falls outside 0.15x to 5.00x. This is a sanity bound, not a judgement about valuation: a method claiming a stock worth 6% of where it trades, or five times it, has almost always consumed a broken input rather than found something.

If that gate rejects every method, the blend does not fall back to the raw numbers. It runs one last absolute filter at 0.05x to 10.0x and keeps whatever survives. On a real distressed name where the per-share denominators are all distorted, the difference matters. Given an EV/EBITDA estimate of 155,100.00, a price-to-sales estimate of 900.00 and a price-to-book estimate of 0.50 on a 100.00 stock, the regular gate rejects all three; the fallback keeps the 900.00 and drops the other two. If nothing survives even that, the blend returns no fair value at all, and the page says so. A stock page showing nothing is better than one showing 36,168.00.

2. The trim, in log space

Among the methods that clear the gate, the blend computes the median fair value and the median absolute deviation around it — both in log space, so a method twice the median and a method half the median are treated as equally far away. Anything more than 3.5 median absolute deviations from that anchor is eligible to be dropped.

Two deliberate limits. The multiplier is lenient by design: for a normal distribution 3.5 MADs is roughly 5.2 standard deviations, so a coherent group of low methods survives intact and only genuine garbage is cut. And the trim never reduces the surviving set below three methods — if it would, it re-expands. A blend built on two methods is not more trustworthy for having thrown a third away.

3. Protection for the methods a sector actually relies on

The trim on its own is sector-blind, and that makes it dangerous. Dividend discount is the anchor for a regulated utility. Funds-from-operations is the anchor for a REIT. Those are exactly the methods most likely to sit apart from the pack — and cutting one because it is the statistical odd one out would flip the answer on precisely the companies where the method is right.

So a method is protected from trimming when it ranks in its sector's top three by weight, or carries a normalised weight of 0.12 or more. Reading the shipped weights, that means:

Financials protects five methods rather than three because five clear the weight floor. That is the rule working as intended, and it is also where the cost lives — see below.

4. Family collapse

The last defence is the one that does most of the work, and it is not an outlier rule at all.

Twenty-five of the methods are mapped into eight families by the input they consume: everything driven off earnings per share in one, off book value in another, off EBITDA in a third, off free cash flow in a fourth. Blending happens in two stages. Within a family, the member methods collapse to a single representative value. Only then do the families vote against each other, each carrying its own combined weight.

The point is correlation. Eight relative multiples that all divide by the same sector multiple are not eight independent opinions, and letting them cast eight votes against a lone discounted cash flow is arithmetic that mistakes repetition for agreement.

What that does to the rogue 480

Back to the eight methods on the 100.00 stock, with the 480.00 arriving as a Graham number. Every figure below came out of the shipped blending function, not a reconstruction of it.

On the default sector profile:

Note what did not happen. The 480.00 was not trimmed — on the default profile the Graham number carries a weight of 0.165, above the 0.12 floor, so it is protected from the trim by rule 3. It went into the blend.

And note what stopped it anyway. The method count rose from seven to eight, but the family count stayed at six. The Graham number landed in the book-value family, which already held price-to-book at 114.00. Instead of casting its own vote at 480.00, it moved that one family's representative value and the family still voted once. A 40.4% error on a plain average became 2.7% on the published midpoint.

Run the identical inputs against the Utilities profile and a different mechanism fires. There the Graham number is not top-three and does not clear the floor, so the trim drops it outright and the blended midpoint is 105.35 either way — the rogue method has no effect whatsoever.

Same bad number, same eight methods, two different defences, depending on sector. That is a design choice, and it is worth being explicit that it is one.

The honest cost

Three things follow that are not flattering, and it is better to state them than to let a reader discover them.

2.7% is not zero. On the default profile the rogue method did move the published number, and moved it in the direction of the error. The margin-of-safety differential — a model estimate, not a prediction — went from 6.93% to 9.41% on a stock where nothing had changed except one broken input. Containment is not immunity.

Protection can protect the wrong thing. Rule 3 exists to stop the trim cutting a sector's real anchor, and the price of that is exactly the case above: when the broken method is the sector's anchor, the rule shields it. There is no version of this trade-off that gets both. We chose to accept a contained error on a high-weight method over a flipped answer on a blue chip whose sector-appropriate method got cut for standing apart.

Family collapse blunts good disagreement too. The mechanism cannot tell a broken Graham number from a correct one that genuinely disagrees. A book-value method that is right about a deeply asset-cheap company gets the same one-vote-per-family treatment. That is the trade for not letting eight correlated multiples out-shout one cash-flow model.

None of these mechanisms decides whether a stock is worth owning, and none of them is a forecast. They govern how much a single broken input is allowed to move an estimate under stated assumptions.

Where this is visible

The method agreement counts on a stock page — how many methods place fair value above the current price and how many below — are computed independently of everything above. They read the stored per-method fair values straight off the row and compare each one to the price, over a fixed set of thirteen methods. A method the gate or the trim removed from the blend still has its own stored value, and still appears in that count.

That is worth knowing before quoting the two figures side by side. The count answers "how many methods came out above the price". The blended midpoint answers "what the weighted, family-collapsed estimate is once the gate and the trim have run". They are different questions over different sets, and on a stock where the trim fired they will not always point the same way.

The full methodology covers the weighting and the stages; the eight-method blend guide walks through what each method does and why they diverge in the first place. Two related pieces on the same theme: nine methods that abstain without net debt, which is the case where the right answer is no number at all, and when reported earnings are not earnings, where the broken input was upstream of every method at once. The Graham number calculator shows the method from this article in isolation.

Frequently Asked Questions

Why not just drop the highest and lowest method every time? A fixed trim discards information whether or not anything is wrong. On a stock where all eight methods agree closely, cutting the extremes changes the answer for no reason; on a stock where one method is genuinely broken, cutting one extreme leaves it if two methods are bad. Anchoring on the median and the dispersion around it means the rule fires in proportion to how unusual a value actually is.

Why measure the outlier distance in log space? Because valuation errors are multiplicative. A method returning twice the median and one returning half the median are equally wrong, but in raw terms the high one looks much further away, so a linear rule would cut high estimates far more often than low ones. Taking logarithms makes the two symmetric.

Does a trimmed method still count toward method agreement? Yes. The agreement counts are read from the stored per-method fair values rather than from the blend's surviving set, so a method the trim removed from the fair value estimate still contributes its own comparison against the price. The two figures answer different questions and are best read together with that in mind.

What happens when only two or three methods have their inputs? The median-and-dispersion trim is skipped, because a median over three values is not robust. A lighter cap applies instead — a single method more than 2.5 times the median of the others is dropped, always keeping at least one — and the fair value estimate is separately compressed toward the current price to reflect the thin coverage.

Can one method still change the published number? Yes, as the 2.7% case above shows. The mechanisms limit how far a single broken input can move an estimate; they do not eliminate its influence, and no blending rule can.

This content is for educational and informational purposes only and does not constitute investment advice. Equity Rank is not a registered investment adviser. The figures above are arithmetic demonstrations produced by calling the blending functions on constructed inputs at a 100.00 price; they describe the mechanism and do not describe any security. All fair value figures are model estimates under stated assumptions, not price targets and not predictions of where any security will trade. Model logic and weights are revised over time.