Screens · Method
Stock Selection and Screen Design
A screen is a set of thresholds applied to a universe. Almost everything that decides whether its output means anything happens before the interesting filter runs, in the definition of the universe, and in when the thresholds were chosen.
The two decisions that matter, and neither is a filter
Screens are usually discussed in terms of their conditions: which indicator, which level. Those are the least consequential part. Two other decisions determine whether the output describes the market or describes the screen.
What was in the universe. Run a screen over "every symbol" and it includes closed-end funds, preferred shares, exchange-traded products, warrants, units and multiple share classes of the same company. Several of those move together on interest-rate news, which means a screen can return a list that looks like a market-wide pattern and is one bond-market move counted forty times.
When the thresholds were chosen. Fixed in advance, a screen is a test. Adjusted until the output looked reasonable, it is a search, the identical mechanism the method page demonstrates with moving-average periods, and just as capable of producing a convincing result from nothing.
The counts themselves are synthetic and the shape is not. Every real screen has this funnel, the interesting filter always runs last on a much-reduced population, and the description of the output is almost always written as though it had run on the market.
Notice also that one filter is not yet a screen. A single momentum condition still returns a list far too long to read, which is why real screens add conditions, and every condition added is another threshold that can be adjusted until the output looks right. That is the tension at the centre of screen design: the second and third filters are what make the list usable, and they are also where a test quietly turns into a search.
The order to build them in
| Stage | Decision | What it protects against |
|---|---|---|
| 1 · Universe | Common stock only, on stated venues, with a stated data source. | Funds and preferreds moving together on rate news, and one company counted three times. |
| 2 · Price floor | A minimum price, five dollars is the conventional choice. | Tick-size artefacts, spreads that are a large fraction of price, and percentage rankings dominated by cheap issues. |
| 3 · Liquidity gate | A dollar-volume floor, measured over a month rather than a day. | Instruments that cannot absorb a position, and single-session volume flukes. |
| 4 · The filter you care about | One condition, with its threshold fixed before you look. | The search. One condition chosen in advance is a test; five conditions tuned together is a story. |
| 5 · A sanity count | Record how many survived each stage, every time you run it. | A screen that has quietly stopped returning anything, or has started returning everything. |
The last row is the cheapest discipline in the whole exercise and the one nobody keeps. A screen whose output count swings from four to four hundred as conditions change is telling you that its thresholds are absolute where they should be relative, and the count is the only place that shows.
Absolute against relative thresholds
A threshold on a bounded measure can be absolute. RSI runs from zero to a hundred by construction, so "above 70" means the same thing in every market, which is why it is the one conventional level this reference does not object to on principle, though it still recommends checking its distribution on your own universe.
A threshold on an unbounded measure cannot. "Volume above two million shares" selects almost nothing in a quiet period and almost everything in a busy one, and the same is true of any dollar, point or range figure. Those filters have to be expressed as percentiles of the current universe, or the screen is partly measuring the market’s overall state rather than the instruments in it.
This is the same distinction that runs through every page in the indicator library: a measure comparable within itself is portable, and a number borrowed from someone else’s sample is not.
What a screen is honestly for
Finding things to look at. That is a modest description and it is the use that survives scrutiny: the market has thousands of instruments, attention is finite, and a mechanical filter is a reasonable way of deciding where to spend it.
What it is not is a source of expected return. A screen with a performance figure attached has almost always been tuned until the figure appeared, and the ways of tuning are numerous enough that it happens without any intention to mislead, one more condition, a slightly different level, a different starting year. The honest version reports the thresholds, the universe, the count at each stage, and no performance claim at all.
Which is why the screens on this site are described rather than run. Each page states what its filter selects and where that selection misleads, the oversold list filling with deteriorating businesses, the most-traded list filling with cheap ones. Those are properties of the arithmetic, they are true on any data, and they are what someone building their own screen needs to know before choosing a threshold.
Frequently asked questions
What does a screen actually do?
It applies a set of thresholds to every instrument in a defined universe and returns those that pass. That is all. There is no judgement in it and no forecast. What makes one screen useful and another useless is not the cleverness of the filters but the two decisions around them: which universe was screened, and whether the thresholds were chosen before or after looking at the results.
Why does the universe matter so much?
Because it decides what can possibly appear. A screen run over every listed symbol includes funds, preferred shares, exchange-traded products and multiple share classes of the same company, none of which is an operating company, and several of which move together for reasons unrelated to equities. Restricting to common stock is the single highest-value decision in screen construction, and it is usually omitted because the data provider does not make it easy.
In what order should the filters run?
Universe definition first, then the liquidity and price gates, then the filter you actually care about. The order does not change the final set — set intersection is commutative — but it changes what you understand about the result: knowing that your interesting filter ran on two thousand instruments rather than six thousand tells you what the output is a sample of. The figure on this page is that count at each stage.
Why exclude low-priced instruments?
Because most indicator arithmetic degrades on them. A stock trading near a dollar moves in large percentage steps for reasons of tick size rather than of information, its spread is a substantial fraction of its price, and it dominates any ranking based on share volume or percentage change. A price floor is a crude filter and it removes a large share of the instruments where a screen produces noise rather than results.
What is the most common mistake?
Adjusting the thresholds until the output looks good. That is a search rather than a test, and it is the same mechanism as the parameter sweep on the technical-analysis page: with enough thresholds to tune, a screen can be made to select almost any historical set. The discipline is to fix the thresholds first, then look, and to count how many variants you tried before reporting any result.
How many results should a screen return?
Enough to be a list and few enough to be read: somewhere between ten and fifty for most purposes. A screen returning three instruments is measuring its own thresholds rather than the market, and one returning four hundred has not filtered anything. If the count swings wildly from day to day, the thresholds are absolute where they should be relative, a percentile of the current universe rather than a fixed number.
Should thresholds be absolute or relative?
Relative, wherever the quantity has a distribution that moves. "RSI above 70" is an absolute threshold on a bounded measure and is defensible; "volume above two million shares" is absolute on an unbounded one and will select nothing in a quiet market and everything in a busy one. Expressing such filters as percentiles of the current universe keeps the size of the output stable and makes the screen comparable with itself over time.
Does a screen need a backtest?
It needs one before any claim is made about what its output does next, and it needs none at all if the screen is being used as what it is, a way of finding things to look at. Most of the value of screening is in the second use, and most of the trouble comes from the first: a screen with a performance figure attached has almost always been tuned until the figure appeared.