Screens · Momentum
Oversold Stocks, What the Filter Selects
An oversold screen ranks instruments by an oscillator at the low end of its scale. It selects on price behaviour and nothing else, which is exactly why businesses in genuine trouble fill the list.
What the filter selects
Almost every oversold list is one line of arithmetic: compute RSI over a fixed period for every instrument in a universe, and keep the ones below 30. Some versions use a stochastic oscillator or a percentage below a moving average instead; the character of the output barely changes.
The word does more work than the formula supports. "Oversold" implies that something has been sold past its worth, and nothing in the calculation knows what anything is worth, no earnings, no balance sheet, no comparison with any other instrument. The reading states one fact: over the lookback window, losses have dwarfed gains. Every other meaning is supplied by the reader.
Why the list looks the way it does
Run the screen for a month and the same character recurs: companies that have missed earnings, sectors under pressure, and a scattering of instruments in genuine distress. This is not a flaw in the implementation. It follows from what persistence means.
An ordinary pullback in a sound business is brief, a few sessions, and the oscillator is back through the middle of its range. To be oversold on the day you happen to run the screen, and again the following week, an instrument has to keep falling, and the most reliable reason for that is that something is genuinely wrong. The filter therefore selects, with no intention of doing so, disproportionately for bad news.
The screen is still worth running. What it is not is a list of bargains, and treating it as one is the single most expensive misuse of a mechanical filter, the more so because the losses arrive slowly, one justified decline at a time.
The three checks that make the list useful
A name from this screen is a candidate for work, not a conclusion. Three questions separate the two, and none needs a subscription.
Why did it fall? An earnings miss, a sector-wide move, an index deletion and an accounting investigation look identical in the closing prices and call for entirely different responses. This takes a minute and eliminates most of the list.
What is the volume doing? A decline on progressively lighter volume is running out of sellers; the same decline on expanding volume is finding new ones. The oscillator cannot see the difference — it reads closing prices only — and it is the more informative of the two readings.
Is everything oversold at once? On a handful of days each year the screen returns hundreds of names, because the whole market fell together. An individual reading in that context says nothing about the individual instrument, and the useful measures are the breadth ones, how many issues took part, rather than anything computed from one price series.
Where it misleads
| Situation | What goes wrong |
|---|---|
| Read as "cheap" | The calculation knows nothing about value. It reports the recent balance of losses to gains and nothing else. |
| Sustained downtrend | The oscillator can sit below 30 for weeks. Buying each new reading means buying repeatedly into a decline. |
| Inherited threshold | Thirty is a convention from 1978, not a property. Derive the level from the instrument's own distribution if it will be acted on. |
| Mixed periods | A list built on RSI 7 and one on RSI 21 are different populations. They cannot be compared or combined. |
| Market-wide selloff | When hundreds of names qualify at once, the individual reading carries almost no information about the individual instrument. |
| Illiquid issues | Without a liquidity floor the list fills with instruments whose closing prices are set by a few trades. |
Frequently asked questions
What does "oversold" actually mean?
That an oscillator has reached the low end of its scale, conventionally RSI below 30. The word implies a judgement the arithmetic does not make: nothing in the calculation knows what the instrument is worth, so it cannot say anything has been sold beyond its value. What the reading states is narrower and true: recent losses have dwarfed recent gains over the lookback window. Whether that is an opportunity or the early part of a justified decline is outside what the number can see.
Why are falling businesses over-represented on the list?
Because the filter selects on price behaviour alone, and the most reliable way for an instrument to be persistently oversold is for the business behind it to be deteriorating. An ordinary pullback in a sound company is brief, so it appears on the screen for a day or two; a company whose earnings are collapsing can sit below 30 for weeks and will therefore appear on almost every list you run. The screen has no way to tell the two apart, because in the closing prices they look identical.
Can a stock stay oversold?
For weeks, and this is the practical fact that ruins the naive use of the screen. In a sustained decline RSI compresses into the lower part of its range and stays there, the figure on this page shows a series holding below 30 for a long stretch while price falls another quarter. Buying each new oversold reading in that kind of market means buying repeatedly into something that keeps going, which is how a mechanical oversold rule loses money faster than almost any other simple system.
Is 30 the right threshold?
It is a convention from Wilder’s original writing, not a statistical property, and its main virtue is that a great many people watch it. Lowering it to 20 produces far fewer names, most of them in genuine distress; raising it to 40 produces a long list of ordinary pullbacks. If a threshold is going to be acted on, derive it from the instrument’s own distribution, the 5th percentile of its own RSI history means something, whereas an inherited number describes whatever was on someone else’s screen.
Which period should the RSI use?
Fourteen by default, because that is what everyone else plots and the coordination is the only real property the number has. A shorter period reaches both extremes several times a month and produces a longer, noisier list; a longer one may not reach either in a quarter. Whichever you choose, the list is only comparable with itself, an "oversold" list built on RSI 7 and one built on RSI 21 are different populations, not different opinions.
Does an oversold reading mean a bounce is due?
No, and the word "due" is where the money is lost. The reading says the recent balance of losses to gains has been extreme; it contains no information about what happens next, and in a strong downtrend the extreme persists. Mean reversion from oversold readings does appear in aggregate over short horizons in some markets and periods, but it is a statistical tendency across many instruments, not a property of the one you are looking at.
What should I check before acting on a name from this list?
Three things, in this order, and all three are free. First, why it fell, an earnings miss, a sector move, an index deletion and a fraud investigation are four completely different situations that look the same in the price. Second, the volume: a decline on progressively lighter volume is running out of sellers, while one on expanding volume is finding new ones. Third, whether the whole market is oversold at the same time, which makes an individual reading much less informative.
How does this differ from an overbought screen?
It is the mirror in arithmetic and not in behaviour. Both extremes are produced by persistence, but the persistence at the low end is more often justified by something real, and declines are faster and more synchronised than advances, which is why oversold lists lengthen dramatically on a few days each year and overbought lists fill in slowly. The asymmetry runs through every measure on this site and it is the reason the two screens should not be read as symmetric opposites.
Is there a live version of this screen?
Not yet. Running it needs a daily price feed, and a stale list of names would read as current to anyone arriving from a search result, which is worse than no list. What this page carries instead is what the filter selects, how it is built, and the checks that turn a name on such a list into something worth an afternoon.