Screeners That Encode a Thesis, Not a Shopping List

The pile of filters that nobody can act on

The usual sequence goes like this. You open a screening tool, and because the interface is a list of checkboxes, you start ticking. Market capitalisation above some floor, because you do not want the very small names. Return on equity above some level, because good businesses earn well. A price to earnings ratio below some ceiling, because you do not want to overpay. Price above the 200-day average, because you would rather be with the trend. Daily volume above some minimum, because you need to be able to get out. Each one is defensible in isolation. Every one of them is a sentence you could say out loud without embarrassment.

Then you press run, and one of two things happens. Either three hundred names come back, which is not a shortlist but a description of the market, or nothing comes back at all. And then comes the step that quietly ruins the exercise: you adjust the thresholds until the count feels comfortable. Twenty names feels right. Fifteen feels better. You nudge the earnings multiple from 25 to 30, the list goes from four names to nineteen, and you carry on.

At that moment the objective changed without anyone announcing it. The thresholds are no longer expressing a belief about what makes an instrument worth owning. They are expressing a preference about how long the output should be. Everything downstream inherits that substitution, and because the screen still looks like a rule, it keeps the appearance of rigour it no longer has.

Here is the test that separates a screen from a pile. Write down, in one sentence, what you believe and why this particular set of filters expresses it. Not a paragraph, not a list of virtues, one sentence with a subject, a mechanism and a horizon. If you cannot write it, the screen has no thesis, and a screen without a thesis produces an arbitrary list. It will still produce a list. That is the problem: the output looks identical either way.

This is upstream of testing, which is worth saying plainly because the two get confused. A great deal of careful writing exists on how a backtest deceives you, including our own catalogue of the eight ways a backtest can lie. All of it assumes you already have a rule worth testing. The screen is where the rule comes from, and a rule assembled by ticking boxes carries a defect that no amount of downstream validation will surface, because the validation will faithfully measure the arbitrary thing you built. If you are still deciding where screening sits inside a wider process, our overview of systematic trading in India places it in the sequence.

One sentence, four claims, four numbers you can defend

A thesis is not a mood. It is a sentence that makes claims, and each claim is either measurable or it is not. The translation from sentence to screen is mechanical once you accept that constraint, and doing it in the open is the fastest way to discover that half of what you believed was not a claim at all.

Take a sentence and break it into the separate assertions it contains. Give each assertion a measurable proxy, which is almost never the assertion itself but the closest observable thing. Then give each proxy a number, and be prepared to say why that number rather than one twenty per cent away from it. Three steps, and the discipline lives in the fact that the chain has to be complete in both directions.

Both directions matter. A claim with no criterion under it is a belief you are carrying without testing, and it will show up later as a discretionary override, because the part of your thesis the screen could not express is exactly the part you will apply by hand. A criterion with no claim above it is worse, because it is doing real work on the output while answering to nothing. Almost every filter you added out of habit falls into that second category.

The translation also tells you what the screen must not contain, which is the half of the exercise nobody does. If your sentence says the market has already begun to reprice something, then a cheapness filter is not a prudent addition, it is a contradiction, and the arithmetic later in this page shows exactly how expensive that contradiction is.

Notice that the three sentences do not merely have different thresholds. They select from different families of criteria entirely, and the second one deliberately inverts a filter the first one relies on. That is what having a thesis means: it makes some criteria mandatory and others forbidden. A screen that draws a little from every available family is usually a screen with no sentence behind it.

The universe, stated so the arithmetic can be checked

Everything from here is computed rather than asserted, on a synthetic universe built for the purpose. Four hundred instruments are distributed across nine synthetic sectors labelled A to I, each sector given its own stated drift, volatility, typical return on capital, typical size and typical earnings multiple. Price behaviour comes from five hundred daily bars driven by a market factor, a sector factor and an instrument-specific term. Earnings multiples are constructed to rise with the trailing move, so that an instrument which has run gets more expensive, as happens in practice. The whole thing is seeded, so it reproduces exactly.

A synthetic universe is used deliberately. It keeps the exercise reproducible, it lets every structural assumption be stated in the open rather than hidden inside a data vendor, and it avoids implying that any real instrument, sector or index would behave in any particular way. Nothing here is a claim about a market. It is a demonstration of what a screen does to a population.

Across that population, the median instrument traded about Rs 8.3 crore a day (illustrative), earned a return on capital of 15.5 per cent, traded at 25.8 times earnings, and had gained 7.2 per cent over six months. Two hundred and eighty-nine of the four hundred were above their 200-day average, so this is a population in a broad advance rather than a decline, which is the condition under which a momentum screen is at its most flattering.

The screen applied to it is the first sentence from the table above, translated exactly as the diagram shows: median daily traded value of at least Rs 5 crore (illustrative), last price above the 200-day average, six-month price change of at least 20 per cent, and return on capital of at least 18 per cent. Four criteria, four numbers, one sentence behind them.

Where the four hundred went

The most useful single picture in screening is the survivor count at each stage, because it is the only view that shows you how violently a list collapses and which criterion did the collapsing.

Liquidity removed 142 names, leaving 258. The structure test removed a further 77, leaving 181. The momentum criterion was the sharpest cut of the four, taking 181 down to 61. The quality test then took 61 to 28. Twenty-eight names out of four hundred, or seven per cent of the universe, is a plausible working shortlist: small enough to look at individually, large enough that no single name dominates.

Then comes the part that matters. Suppose you look at that list and notice it contains some richly valued instruments, and you add a fifth criterion: an earnings multiple of no more than 30 times. It sounds like prudence. It is the kind of addition that feels like tightening a good screen rather than changing it.

It took 28 names to one. Tighten the same criterion by five multiple points, to 25 times, and the list is empty.

What makes this instructive is that the fifth criterion is not a demanding one. Two hundred and fifty-four of the four hundred instruments in the universe, nearly two thirds, traded below 30 times earnings. On the population it is a permissive filter. Applied to this particular shortlist it removed 27 of 28, because the shortlist was selected for having already run, and instruments that have already run are precisely the ones whose multiples have expanded. The cheapest survivor traded at 27.4 times against a universe median of 25.8.

So the fifth criterion did not refine the thesis. It contradicted it. The sentence said the market has started paying up and will keep paying up; the fifth criterion said do not buy anything the market has started paying up for. That contradiction is invisible when you read the criteria as a list of virtues, and it is unmissable the moment you compute the funnel.

The funnel is a story. The arithmetic is a set.

Having praised the funnel picture, it is only honest to say what it cannot tell you, because two computations undercut the way most people read it.

The first is that the order of the filters does not matter. Running all twenty-four possible orderings of the four criteria produced the same terminal count every time: twenty-eight. That has to be true, because a chain of filters is an intersection of sets and intersection does not care about sequence. But the intermediate counts vary enormously. Depending on the order, the count after two criteria ranged from 40 to 181. So the shape of the funnel, the thing people screenshot and reason about, is a narrative choice. Reorder the same four filters and you get a completely different story about which criterion was doing the work, with an identical answer at the end.

The second computation is sharper. Take each criterion out in turn, run the other three, and see how many names that criterion was uniquely responsible for removing.

The structure test removed nothing. Not a small number, zero. Run liquidity, momentum and quality without it and you get exactly the same twenty-eight names. The reason is mechanical: every one of the 96 instruments in the universe that gained at least 20 per cent over six months was already trading above its 200-day average. The momentum criterion had silently absorbed the structure criterion.

And yet the funnel shows the structure test cutting 77 names, which looks like meaningful work. Both facts are true. It removed 77 names at its position in the sequence, and it removed none from the finished list, because everything it removed would have been removed anyway. The funnel measures the order you happened to choose. The leave-one-out test measures the criterion.

This matters beyond tidiness. A trader who believes their screen has four independent constraints has a false picture of how selective it is, will defend a threshold that does nothing, and will be surprised when loosening it changes nothing. Momentum uniquely removed 36 names and quality uniquely removed 33, so those two are carrying the screen. Liquidity uniquely removed 12. Structure removed none. The test costs four re-runs and it is the single highest-value diagnostic in this page.

Two ways to fail, both measurable before you trade

A screen fails in two directions and the failures look nothing alike.

The loose failure first. Take three gentle criteria: traded value of at least Rs 1 crore a day (illustrative), price above the 200-day average, and a six-month change that is not negative. Every one of those is a sentence a reasonable person would nod at. Run them and 249 of the 400 instruments survive, which is 62 per cent of the universe, spread across all nine sectors. That is not a shortlist. It is the market with the weakest names removed, and no process can act on it consistently, because the human being at the end will pick from it using criteria that were never written down. A loose screen does not remove judgement, it relocates judgement to the least examined place in the process.

The tight failure is the mirror image. Add to the four working criteria a valuation ceiling at 30 times, a debt-to-equity limit of 0.35, and a requirement to be within eight per cent of the 52-week high. Seven criteria, all sensible-sounding. The output is empty. Not small, empty.

The trap sits between them, and it is the reason both failures persist. Faced with an empty list, almost everyone loosens the last threshold they added until something appears. Faced with a list of 249, almost everyone tightens until the count looks manageable. Both moves let the length of the output choose the numbers, which is the original sin restated. The fix is to decide, before running anything, how many names you can genuinely act on, derived from your capital, your risk per position and the hours you actually have. Then, if the output misses that range badly, treat it as evidence that the sentence needs rewriting rather than evidence that a number needs nudging.

The characteristic you did not know you were buying

A shortlist of 28 names looks diversified. Count them by sector and it stops looking that way.

Twelve of the twenty-eight came from a single synthetic sector that made up 13 per cent of the universe. Four of the nine sectors contributed nothing at all. The concentration index of the surviving list was 0.270 against 0.114 for the universe, which is to say the list was roughly two and a half times as concentrated as the population it was drawn from. Nothing in the screen mentioned a sector. The thresholds selected one.

The reason is that screening criteria are rarely independent. In this universe, one sector was constructed to have both the strongest drift and a high typical return on capital, so a screen asking for recent strength and good returns on capital was, without saying so, asking for that sector. Any real screen has the same property, because sector membership is the strongest common driver of both fundamental characteristics and price behaviour.

The consequence is measurable in co-movement. The average pairwise correlation among the 28 survivors was 0.324, against 0.290 for randomly chosen baskets of the same size, and only 92 of 2,000 random draws were as correlated or more. Held equally weighted, the surviving list had an annualised volatility of 16.4 per cent against 15.0 per cent for random baskets of 28, and only 28 of 2,000 random baskets were riskier.

Two honest qualifications belong here, because the tidy version of this argument overstates the case. First, part of that extra risk is not co-movement at all: the survivors were individually more volatile than average, 27.4 per cent against 26.6 per cent. Strip that out by comparing the ratio of average single-name volatility to basket volatility, and the screen list scored 1.67 against 1.77 for random baskets. So the concentration is real and it costs you diversification, but it explains roughly two thirds of the gap rather than all of it. Second, the co-movement effect clears about ninety-five per cent of random draws of the same size, which is a clear signal rather than an overwhelming one.

The cleanest single number is the factor decomposition. Regress the equally weighted basket on just two series, the market factor and the one sector's factor, and 89.8 per cent of its daily variance is accounted for, against 82.1 per cent for random baskets of the same size. Four of two thousand random draws matched it. A trader holding those 28 names owns, in substance, two bets: the market, and one sector. They would describe themselves as holding a diversified list of high-quality names in an uptrend.

The remedy is not complicated. Count the survivors by sector every time you run the screen and put that count in the output next to the names. If one sector is more than a third of the list, you have taken a sector view, and you should either take it deliberately, with the position sizing that implies, or add a constraint that caps it.

A threshold you cannot move is not a threshold

Every number in a screen was chosen by somebody, and most of them were chosen because they are round. Eighteen per cent, twenty per cent, Rs 5 crore. The question worth asking is whether the list would survive a small change to any of them, and the way to answer it is to move each threshold by the same relative amount and measure how much of the surviving list changes.

The results are not symmetric, and the asymmetry is the lesson. Move the six-month price change threshold anywhere in a band twenty per cent either side of its set value, and no more than 10.7 per cent of the list changes. Between 18 and 20 per cent the surviving list is literally identical, which means the chosen number sits in the middle of a plateau. Liquidity behaves similarly: across the full band the worst churn was 10.7 per cent.

The return-on-capital threshold behaves entirely differently. Raising it from 18 to 18.9 per cent, a move of five per cent in relative terms, changed 28.6 per cent of the list and took the count from 28 to 20. Across the same twenty per cent band the churn reached 46.4 per cent. In this universe the distribution of returns on capital piles up right where the threshold was set, so the cut is slicing through the densest part of the population. There is no plateau. The number is standing on a cliff.

Which means close to three names in ten on that shortlist are there because someone typed 18 rather than 19. That is not a thesis being expressed, it is a boundary accident, and it is worth noticing that the unstable threshold was the one that felt most precise. The momentum criterion, the loose-sounding one, was the robust one. Intuition gets this backwards reliably.

Three responses are available and all are better than leaving it alone. Widen the criterion until it sits on a plateau, accepting a longer list. Replace the hard cut with a rank, taking the top names by return on capital rather than everything above a line, which removes the boundary entirely. Or keep the cut and report the plateau width alongside the list, so that anyone reading the output knows how much of it is structural and how much is an artefact of one digit. The same logic governs every tunable number in a systematic process, and our companion page on parameter sensitivity works through what it means when a whole strategy behaves this way.

The five families, and what each one is actually for

Screening criteria fall into a small number of families. Knowing which family you are drawing from is useful mainly as a check on the sentence: a thesis is a claim about one mechanism, so a screen pulling evenly from all five families is usually a screen with no mechanism in mind. The material on which specific ratios are worth using, and how to read them honestly, is covered in our guide to fundamental analysis for Indian retail investors, and is not repeated here.

Liquidity is the criterion Indian conditions punish hardest

Of the five families, liquidity is the one where a naive threshold does the most damage, because traded value is a proxy for something else and the gap between the proxy and the thing widens exactly when it matters.

What you actually care about is impact cost: what it costs you to consume enough of the resting book to fill your order. That is a function of your size against the depth available, not a property of the instrument alone. A name that is comfortably liquid for an order of one lakh may not be liquid for ten, and the threshold that protected you at one size silently stops protecting you as the account grows. The honest form of the criterion is therefore relative: your intended order as a share of the instrument's typical daily turnover, with a ceiling you have chosen. An absolute rupee threshold is a crude stand-in for that, and it should be recalculated whenever your position size changes. The mechanics of reading depth are covered separately in our page on market depth.

Two structural features of Indian equity markets make the point sharper. Exchanges apply daily price bands, so an instrument can be locked at a limit and effectively untradeable at any price for the session, which no historical average of traded value will warn you about; the mechanics are set out in our page on circuit limits. And in the derivatives segment, positions are expressed in fixed lot sizes, so the smallest position you can take is not one unit but one lot, which means a liquidity filter and a position-size floor interact: an instrument can pass your liquidity test and still be untradeable for you because a single lot breaches your risk budget. That interaction is worth understanding before it bites, and our page on lot sizes explains how it is set.

One more practical point. Traded value is not stable. A single event day can lift an instrument's recent average enough to clear a threshold it would otherwise miss, which is why a median is a better input than a mean, and why a filter measured over a longer window is harder to fool than one measured over a week.

What a screener cannot do

Four limits are worth stating explicitly, because a screen that is honest about them is more useful than one that pretends to be a decision engine.

It ranks and filters. It does not decide. Nothing in a screen output tells you to buy anything. It tells you which instruments currently satisfy a set of conditions, which is a much narrower statement than it feels like when you are looking at a tidy list of names. Everything between the list and the position, the entry, the size, the exit and the abandonment rule, has to exist somewhere else.

It cannot tell you whether the characteristic it selects for is currently being rewarded. This is the deepest limitation and the least discussed. A momentum screen returns momentum names in every environment, including the environment that is punishing them, because the screen is a description of a characteristic and carries no information about whether that characteristic is in favour. The list looks equally confident either way. Establishing whether a characteristic has been rewarded, and over what horizon, is a separate exercise requiring the full apparatus of a properly run backtest, and it has to be done on the characteristic rather than on the screen output.

It is a snapshot with no memory. Run the same screen next month and it will hand you a list assembled from scratch, with no knowledge of what it told you last time. Left unmanaged, that generates turnover you never chose: a name that drops one place below a threshold disappears from the list, and you sell something for a reason nobody wrote down. A screen intended to drive positions needs an explicit rule about what happens to a holding that falls out of the list.

It inherits every property of its universe. If the instrument list you screen is today's constituents of an index, then every name in it survived to today, and that bias passes straight through into your output regardless of how careful the criteria are. The screen cannot see the population it was never given.

Before you trade off a screen

The checks below are the ones that repeatedly change a list, and none of them requires anything more than re-running the screen with one input altered. Two of them, the leave-one-out and the threshold sweep, are the diagnostics that produced the two most surprising results on this page. If a screen is going to feed real positions, it also has to connect to the machinery around it, which is the subject of our guide to building a trading system, and to a written plan, for which our trading plan generator is the practical starting point.

A screen is a belief you can run

The reason to treat screening this seriously is not that the screen is the important part. It is that the screen is the first place where a vague view becomes a specific, executable, checkable object, and everything that is wrong with the view becomes visible the moment you try to write it as criteria. The exercise on this page found three defects in a screen that looked entirely reasonable on paper: a criterion that did nothing, a sector bet nobody had chosen, and a threshold that was moving a third of the list for the sake of one digit. None of those were discoverable by reading the criteria. All three fell out of arithmetic that took minutes.

This is also the point at which a discretionary trader becomes a systematic one, which is a smaller step than it sounds. Nothing above requires a model, a data subscription or a language you do not already know. It requires writing one sentence before you write any filters, and then being willing to run three diagnostics that might tell you your screen is not what you thought. The people who never make the transition are usually not blocked by technique. They are blocked by not wanting to know.

What you get in return is a list you can defend. Not a list that will be right, because no screen is right, but one where you can say why every name is on it, what would take it off, and which of your beliefs would have to be wrong for the whole list to be worthless. That is a different kind of object from a pile of filters, and building it well is a large part of the method we teach.

Every count comes from a single deterministic simulation, seeded so that it reproduces identically on each run. Four hundred synthetic instruments are allocated across nine synthetic sectors, each with a stated drift, volatility, typical return on capital, typical size and typical earnings multiple. Five hundred daily bars are generated from a market factor, a sector factor and an instrument-specific term; the market and sector shocks are de-meaned inside each measurement window so that each sector's realised drift equals its stated drift, which keeps the demonstration from depending on a lucky or unlucky draw. Fundamentals are drawn around the stated sector levels, and earnings multiples are constructed to rise with the trailing move. Survivor counts, leave-one-out contributions, sector concentration, correlations and threshold churn are all computed directly from that one population, and the two thousand random comparison baskets are drawn from the same universe.

All results are illustrative and simulated. They are not a track record, they are not a forecast, and they describe no real instrument, company, sector or index. Rupee figures are illustrative parameters of the simulation, not observations. The purpose of the exercise is to demonstrate properties of the screening procedure itself, which is why the universe is synthetic and fully specified rather than borrowed from a live market. Nothing on this page is a recommendation to buy, sell or hold anything.

• Walk-forward analysis, demonstrated →
• Measuring whether an edge exists →
• What a breakout actually is →
• Paper trading and what it can prove →