Top-Down vs Bottom-Up Research: Where the Work Starts

Top-down research starts from the economy or a field and narrows to a company. Bottom-up research starts from a company and reaches the field only afterwards, as context. Each is strong precisely where the other is weak, and each produces one characteristic error. The useful question is therefore not which route to use, but which error the starting point has exposed the analyst to.

Underneath that sits a fact about evidence rather than a fact about method. A set of accounts records what happened to one company and never records why, and a field aggregate records what happened to everybody and never records who caused it. Neither body of evidence contains the other, so wherever the reading begins, there is a class of explanation that is not yet visible. The invisible class can be named exactly, and it is different for each starting point.

What is top-down research, taken on its own?

Top-down research begins with conditions that apply to many companies at once and works inward until it reaches one. The analyst establishes what is happening to input prices, to demand, to capacity and to the cycle, then asks which fields those conditions favour or squeeze. Only after that are the accounts of anything opened, and by then the ordinary shape of a year is already settled before the first line of a filing is read, so the filing is being measured against something instead of read in a vacuum.

The defining feature of top-down research is that the company is the last thing examined and is chosen by the conditions, rather than the conditions being fetched later to explain a company already chosen. That ordering is not a stylistic preference. The ordering decides what evidence is available at the moment each judgement is made, and a judgement made before the relevant evidence arrives is not improved by the evidence arriving afterwards.

The everyday version: a woman is deciding what to sell at a temple fair. She could pick a stall type she likes and then find out how the fair goes. Or she could first walk the ground on the two previous mornings, count how many people arrive, notice that everyone is carrying water in the heat and that there are nine snack stalls and one cold drinks stall, and only then decide what to sell. The second walk is top-down. The walk does not tell her she will run a good stall. The walk tells her which stall the conditions are pushing toward. The answer is genuinely useful and genuinely incomplete.

What is bottom-up research, taken on its own?

Bottom-up research begins with one company and refuses to look anywhere else until that company has been read properly. The questions are about this business and no other: the products that carry the volume, the ones that earn anything, what a unit absorbs before it leaves the plant, the customers who settle on time and the ones who do not, and where three years of cash actually went. The statements are read as one continuous object across several years rather than as a table of ratios. The field enters late, and it enters as context for something already found rather than as the frame it was found in.

The defining feature of bottom-up research is that all the evidence it collects belongs to one identified business rather than to a category average. One identified business is exactly what produces specific knowledge, and exactly what leaves that knowledge with nothing to measure it against. A company that improves by three points has improved by three points. Whether three points is a lot depends entirely on something the reading has not yet fetched.

The everyday version again. A tailor hands over one month of her shop's books and asks whether the shop is doing well. The books can be read thoroughly. Fabric cost, rent, how many blouses, what she charges, how long customers take to pay. At the end the reader knows that shop better than the tailor does, and still does not know whether the good month was hers or whether every tailor on that road had a good month because a wedding season landed early. Nothing in her books contains the other tailors.

What is each route genuinely good at?

The field first route is good at noticing that a whole field is moving. Noticing a whole field move sounds modest, but a field wide movement is close to invisible from inside a single company. In the latest year the coatings field grew from Rs 43,500 crore to Rs 48,300 crore, a rise of 11.0 per cent. A set of company accounts contains one company's revenue and nothing else, so a company first reader who never fetches those two field numbers has no way to produce them.

The company first route is good at understanding what a specific business actually does. A field aggregate cannot show that a maker sells into two quite different channels on quite different terms, that one of them prices annually and the other prices on contract, or that a new line is sitting in capital work in progress and has not yet earned anything. Channel, pricing basis and idle capacity are readable in one place only, the company itself. A reader who works only from field data will get them wrong at a fairly steady rate.

Breadth and depth are two different capabilities rather than two styles of the same capability. A researcher cannot compensate for lacking one by being better at the other. A very good company first reader does not slowly become able to see the field. A very good field first reader does not slowly become able to see inside a business. What is missing in each case is evidence rather than skill, so the two abilities never converge with practice.

What does each route systematically miss?

The field first route misses which participant actually captures a favourable condition. Establishing that input prices moved and that realisation ran ahead of them shows that a gain existed. The same finding does not show where along the chain the gain settled. A move of that kind passes through a supplier, then through a maker, then through a dealer, and each of those can keep some, pass some on or lose some. Narrowing from the field to a maker on the grounds that the maker is in the field selects a company for its membership. Membership is not evidence about the company at all.

The company first route misses that a company achievement is a shared condition. The route reads a real improvement, records it accurately, and attributes it to the only thing in view. A shared condition and a company specific one leave identical marks in a single set of statements, so there is no moment in the process at which the accounts object. Most people expect more care to be the answer here. More care is not the answer.

Both blind spots are systematic rather than occasional. Neither can be closed by being more careful inside the route that has it. A blind spot that appears sometimes is a lapse and responds to discipline. A blind spot built into the shape of the evidence appears every time, in exactly the same place, and responds only to fetching evidence from outside the route. The distinction between a lapse and a built in gap is the difference between a habit worth improving and a structure worth changing.

Why is neither route a complete process on its own?

Put the two blind spots next to each other and the conclusion arrives without any further argument. The field first route cannot say which participant captured a movement. The company first route cannot say whether a company finding was shared. Each of those questions is answered by evidence that the other route routinely collects. So a process built on one route alone has a gap that is known in advance, sits in the same place every time, and is not closed by any amount of work inside the route.

Two blind spots side by side make a claim about coverage rather than a recommendation to compromise between two styles. Nobody is being asked to be a bit of each and good at neither. The point is narrower and harder. A single route process can be run superbly and still be missing a fixed, nameable class of explanation. Nothing in the analyst's own work is going wrong, so the absence is never felt.

The difference is audible in how the two failures get described afterwards. A process that is merely sloppy produces mistakes that look like mistakes: a figure transposed, a period compared against the wrong one, an assumption never stated. A single route process produces work that looks finished. The sentence is well supported, the arithmetic ties, and the evidence that would have changed the conclusion was never in the room to be checked.

How do the two routes get combined in practice?

The combination is an order rather than an average. Field first, to establish what conditions everybody was under. Then company, to establish what this particular business did with those conditions. Then back to the field, to test whether the company finding is actually a company finding or whether the neighbours have the same one. Three steps, and they are not interchangeable: each one supplies the benchmark that the next one needs.

The third step is the one people skip, and it is the step that catches the characteristic company first error. Skipping it leaves the whole sequence exposed to the very failure the sequence was built to prevent. It is skipped for an understandable reason. By step two there is a finding, the finding is interesting, and going back out to check two competitors feels like retracing ground already covered. The return is not retracing. Steps one and three ask different questions of the same field: step one asks what the conditions were, and step three asks whether the company specific finding shows up in the neighbours too.

Notice that the third step is cheap. The step is not a second research project. For the case worked below it is one column of gross marginWhat is left of a rupee of sales once the direct cost of producing the goods has been taken out, shown as a percentage. The accounting material builds it. movements for two competitors over the same period, which is a single afternoon of retrieval against a finding that would otherwise carry a report.

Both routes worked on the same company, to different answers

Sarvani Coatings Limited, an invented maker of decorative and industrial coatings, is listed and reports the figures below. Take it through both routes and watch where each one stops.

Run the field first route. Field revenue rose from Rs 43,500 crore to Rs 48,300 crore in the latest year, a rise of 11.0 per cent. Inside that total Nandivarman Paints Limited is Rs 14,490 crore or 30.0 per cent, Sarvani Coatings is Rs 2,415 crore or 5.00 per cent and Kesaria Surface Solutions Limited is Rs 1,449 crore or 3.0 per cent, so the three named makers together are 38.0 per cent of the field and the other 62.0 per cent is scattered across many smaller makers. The dominant input is a short list of pigments and resins, and their prices all move on one common driver, so a cost movement lands across the makers inside much the same window. Over the two published years realisationRevenue divided by units sold, so the average price actually achieved per unit rather than the list price. The earnings material builds and uses it. ran ahead of input cost across the whole field. All three named makers gained gross margin over those two years. So the conditions were favourable, and the route now narrows to a maker inside the field. Sarvani Coatings is a maker inside the field, so Sarvani Coatings is selected.

The characteristic field first error is sitting in that last sentence: the company was chosen for its membership of the field rather than for anything about the company, and the participant who may have captured most of the input move was never in the search at all. Thottam Chemicals Limited sells resins and additives to the makers and sits one step up the chain. An input price move arrives at the makers through Thottam Chemicals. Whether Thottam Chemicals passed the whole of that move on or kept some of it is a real question with a real answer, and the answer is not in the record: Thottam Chemicals' own margin was never collected. The certain part is narrower and more damaging. A route that goes from the coatings field to the makers in it and then to one maker never asks the question, so it cannot get it wrong and it cannot get it right.

Now the company first route runs on the same company. Read alone, Sarvani Coatings has genuinely good accounts. Gross margin stood at 43.0 per cent in year one and 46.0 per cent in year three, so the two year move is 3.0 points. In the latest year on its own, revenue was up 13.9 per cent and gross profit climbed from Rs 933 crore to Rs 1,111 crore, a rise of 19.1 per cent. Earnings before interest, tax, depreciation and amortisation (EBITDA) added 31.2 per cent and profit after tax added 41.1 per cent. Every step further down the ladder carries a bigger percentage than the step above it. Margin expansion looks like that once it reaches the bottom line. A reader who stops here records an improving company and has recorded nothing false.

The table is a one year window, year two into year three, and the 3.0 point margin gain quoted above is a two year window, year one into year three. Both are true and they are not the same measurement. The period is therefore written beside each of them. Inside the one year window, volume added 6.0 per cent against revenue at 13.9 per cent. Realisation per unit is therefore roughly 7.5 per cent higher. The materials bill for each unit made was about 3.6 per cent heavier over that same year. Realisation and materials together carried the materials line down as a share of revenue, 56.0 per cent to 54.0 per cent. Input cost per unit did not fall. The margin gain is entirely that realisation outran input cost.

The characteristic company first error is sitting in that reading too: across those same two years the gross margin line reads plus 2.4 points at Nandivarman Paints Limited and plus 3.6 at Kesaria Surface Solutions Limited, so every maker in the peer setThe small group of comparable companies a reading is measured against. Choosing one, and defending the choice, is settled in the valuation material. improved and Sarvani Coatings sits in the middle of them. A spread of 1.2 points separates the best and the worst, and Sarvani Coatings is inside it. The peer spread does not make the company's own pricing irrelevant. The spread makes the accounts, on their own, unable to say whether there was any.

Now run the third step. Neither single route contains it. Go back to the field holding both company findings and test each one. The margin gain does not survive: every maker gained, and Sarvani Coatings' 3.0 points sits between 2.4 and 3.6. The growth gap does survive: revenue growth of 13.9 per cent at Sarvani Coatings sits against field growth of 11.0 per cent, leaving 2.9 percentage points of daylight, and if Sarvani Coatings is lifted out of the field altogether what remains went Rs 41,380 crore to Rs 45,885 crore, or 10.9 per cent. Its share of the fieldThe slice of a whole field's sales that one company accounts for. The worth of a movement in that slice is worked out separately. reads 4.87 per cent in the earlier year and 5.00 per cent in the later one, which is 0.13 percentage pointsThe plain difference between two percentages. Going 4.87 to 5.00 is 0.13 percentage points of movement, a much smaller thing than a 13 per cent rise in anything. of movement. On volume it grew 6.0 per cent against the field at 4.5 per cent, a gap of 1.5 points.

Two findings entered the third step looking identical and left it separated: the margin gain turned out to be shared and the growth gap turned out not to be. The field first route never opened the accounts closely enough to have the two findings, and the company first route never fetched the neighbours it needed to test them, so neither single route produces the separation.

Which error belongs to which starting point?

A diagnosis can be run on a finished reading in about a minute. The question is where the reading started. If it started from the field, the exposure is picking the wrong beneficiaryWhoever actually ends up keeping a favourable movement. A price move can be kept by a supplier, a maker or a dealer, and only one of them is the beneficiary. of a real change: the change was real, the reading of it was right, and the participant selected may not be the one who kept it. If it started from the company, the exposure is crediting a shared condition to one company: the finding is accurate about that company and wrong about what it means.

Knowing which route a reading started from names which of the two errors it is exposed to. A general instruction to be careful becomes a specific thing to go and check. A field first reading gets a cross-checkFetching evidence from outside the route being run, specifically to test a finding the route cannot test on its own. one step up and one step down the chain: who supplies this field, who buys from it, and does the movement found show up more strongly in one of them. A company first reading gets a cross-check sideways: two competitors, the same period, the same line.

Neither cross-check is expensive, and that is the part worth holding on to. The failure described here is not caused by a shortage of effort. The cause is effort pointed in a direction that cannot reach the answer. The fix is therefore a small amount of work in a different direction rather than a large amount more in the same one.