We tested 637 oversold stocks. The bounce is real. The edge isn't.

Start here: what “oversold” actually measures

Almost every screen that prints the word uses the same number, the 14-day Relative Strength Index. It works like this. Take the last fourteen days. Add up the points gained on the up days and the points lost on the down days. RSI is the share of the total movement that was upward, rescaled to run from 0 to 100. Fifty means the up days and the down days cancelled out. Below 30 is the conventional “oversold” line; above 70 is “overbought”.

Read that definition again and notice what is not in it. There is no revenue, no profit, no debt, no valuation, no news. RSI knows only the shape of the last fourteen closes. It is a measure of how one-sided recent selling has been — nothing more. A stock can be oversold and expensive. A stock can be cheap and not oversold at all.

Monday’s list, and what it has in common

Here is everything in the set that closed oversold, ordered by how oversold, with the fall from its own one-year closing high alongside.

Six of the eleven are not company stories at all. Utilities, property, twenty-year Treasuries, junk bonds, and two banks plus the whole financial sector move together because they are all priced off the same input: the interest rate. We will come back to that, because it is the only part of this list with a mechanism behind it.

The test: 637 episodes, five years, one rule

For each of the 55 single stocks we walked the whole five years of closes and marked the first day RSI(14) fell below 30 — first day only, so a name that stays oversold for a fortnight counts once rather than fourteen times, which would otherwise let one bad stretch dominate the result. That gave 637 episodes. From each one we measured the stock’s return over the next 21, 63 and 126 trading sessions — roughly one, three and six months — and the S&P 500’s return over exactly the same days.

The first half of the answer: yes, they bounced. Median return after one month, +2.26%. After three months, +5.38%, with 67.2% of episodes higher than where they started. After six months, +6.03%. If your question was “do beaten-down stocks recover,” the five-year record says usually, and reasonably soon.

The second half is the problem. Over those same three months the S&P 500 rose too, and it rose by more often enough that only 47.5% of oversold entries beat it. Stretch to six months and the figure falls to 39.6%, with a median shortfall of 5.47 percentage points. Six times in ten, the money would have done better sitting in the index.

And the control that decides whether that number means anything: we ran the identical test on every fifth day in the same 55 names over the same window — no signal, no condition, 11,384 entries. Those days beat the index 48.7% of the time at one month and 46.6% at six months.

The bounce is real. It is just the market’s bounce, arriving a little early and then handing back more than it brought.

Two sub-tests are worth naming because both came out against the conventional advice, and we are reporting them rather than quietly dropping them.

• “Only buy the dip in an uptrend” did worse, not better. Splitting the episodes by whether the stock was still above its own 200-day average at the time: the 69 measurable “dip in an uptrend” entries trailed the index by a median 8.22 points over six months and beat it 37.7% of the time. The 509 “oversold inside a downtrend” entries trailed by 5.21 points and beat it 39.9%. The sample is lopsided — oversold mostly happens in downtrends, which is why there are seven times as many of them — so treat this as a caution against the rule, not a reversal of it.
• More oversold was not better. The 37 measurable episodes with RSI under 25 — the really violent ones — trailed the index by a median 9.70 points over six months, worse than the full set. At one month their median return was negative, −0.28%.

The signal that did pay, and it was never about a stock

Run the same rule on the index itself and the picture inverts. The S&P 500 closed oversold on just eleven days in five years. Ten of those eleven were higher six months later. The median gain was +19.26%. The single failure was 21 January 2022 — the opening days of that year’s bear market — and it cost 9.7%.

Eleven observations is a small sample and one of them lost 9.7%, so this is evidence rather than proof. But the direction is not subtle, and it has a mechanism that the single-stock version lacks. When one stock is oversold, the market is telling you something specific about that company, and the market is often right. When the whole index is oversold, several hundred companies have sold off at once, which is far more likely to be about liquidity, leverage or fear than about several hundred simultaneous business problems. Indiscriminate selling is the thing that reliably reverses. Discriminating selling is not.

Which raises the obvious question about Monday. The index is not oversold. SPY’s RSI closed at 50.6 — the exact middle — and the Nasdaq proxy QQQ at 58.5.

What is unusual is the spread underneath. 10.9% of the 55 single stocks were oversold at Monday’s close — the joint-highest month-end reading in the three years charted, matched only by October 2024. An index at dead-centre with a tenth of its members oversold is a market pulling apart, which is the same thing we found in the calm-index, stressed-bond split and again in the AI names nobody quotes.

The mechanism behind “once the rally resumes”

A rally needs a cause. The clearest candidate on the table is the one that has been doing the damage: the 10-year real yield, which is the Treasury yield after subtracting expected inflation — the return lenders demand in genuine purchasing power. It closed at 2.85% on 24 September 2026, up 1.07 points in a year and 0.62 points in three months, and that is the highest it has been in the five years we pulled.

Crucially, almost none of that came from inflation expectations, which fell 0.04 points over the year to 2.34%. The rise is the real rate itself. That matters because the real rate is the number every future cash flow is discounted at, so when it goes up, the assets that promise most of their money furthest away fall hardest — long bonds, utilities, property, and growth stocks whose value sits in profits a decade out.

So we measured it directly rather than assuming it. For each name we regressed its daily percentage move on the daily change in that real yield across the last 248 sessions. The result reads as: this is how much the stock moved, on average, on a day the real yield rose a tenth of a point. As a check that the method works, TLT — a fund that holds nothing but long Treasuries, whose sensitivity to rates is arithmetic rather than sentiment — comes out at −1.37% with an R² of 0.63, by far the tightest fit in the set.

Now put the two lists side by side, because this is the finding that makes the premise of the question difficult. They barely overlap. Across the 55 stocks, the correlation between rate sensitivity and RSI is −0.09 — effectively zero. The names that would gain most from falling real yields are Intel, Micron, AMD, Lam Research, Qualcomm and Applied Materials, and every one of them closed Monday with an RSI between 54 and 66, sitting 11% to 66% above its own 200-day average. Meanwhile the most oversold names in the set — McDonald’s at −0.38%, PepsiCo at −0.44%, Morgan Stanley at −0.35%, Bank of America at +0.01% — barely respond to the real yield at all.

The three exceptions are the interesting part of the list: Boeing (RSI 27.4, −1.62% per tenth of a point, 26.9% below its high), Home Depot (RSI 29.6, −1.47%, 28.7% below) and Oracle (RSI 37.3, −2.03%, 57.6% below its high and 19.2% under its 200-day average). Those are the only three places in this set where “beaten down” and “would benefit most if rates fall” are the same name. Oracle is also the bridge to the previous post: it is the single worst-damaged name in the AI build-out, and its high rate sensitivity is exactly why.

What this means for your money

The bounce you are hoping for is mostly the market’s bounce. Over five years, an oversold large-cap rose about two-thirds of the time within three months, and over six months it beat the index less often than a day chosen at random from the same stocks. The gain was real; the advantage was not. Anyone comparing a single-stock recovery idea against simply holding an index fund is comparing it against something that, on this evidence, won more often.

The condition with the strong record is not present. The version of this signal that paid a median 19.3% over six months fired when the whole index was oversold, eleven days in five years. On Monday the index sat at 50.6. Whatever is happening in McDonald’s and Home Depot, it is not that.

Six of Monday’s eleven are the same trade. Utilities, property, long Treasuries, junk debt and the banks are all priced off the interest rate. Owning several of them is one position in the rate, held in five tickers — which looks like diversification on a statement and behaves like concentration on a bad day.

The cheap names and the rate-sensitive names are different names. Correlation −0.09. If your reason for holding something beaten down is that a rate cut will lift it, check that reasoning against the specific name: McDonald’s moved −0.38% on an average tenth-of-a-point day, and Intel −2.41%. Only Boeing, Home Depot and Oracle are in both camps here.

Rates have not turned yet. The real yield is at a five-year high and still rising — up 0.62 points in three months. Credit spreads, quiet all year, have widened from 2.78 to 2.93 points over the same three months. Both are small moves and neither is a crisis, but they are pointing the same direction, and it is not the direction a rally needs. For how the model reads regime conditions like these, see the methodology.

What would change this read

• The window contains one rate regime. September 2021 to September 2026 covers a historic rise in real yields from −1.17% to +2.85%. In a five-year stretch of falling real rates, the stocks that fall hardest are a different set and the six-month numbers could easily invert. One window is not a law.
• 55 stocks we chose is not the market. They are large, liquid, mostly profitable American companies. Small caps, loss-making growth and anything that went bankrupt in the window are absent, and that last omission flatters every recovery statistic here.
• The comparison is against an index that had an exceptional run. Losing to the S&P 500 between 2021 and 2026 is a low bar to fail. Against a flat index the same episodes would look far better, and the honest version of the finding is “oversold did not beat this index over thiswindow,” not “oversold never works.”
• RSI(14) and 30 are conventions. A different lookback or threshold gives different episode counts and different averages. We report one definition, stated plainly, so you can see what was measured — not because it is the correct one.
• A median is not your outcome. The spread around these numbers is enormous — the 637 episodes range from −46.3% (Tesla, 7 October 2022) to +108.0% (Intel, 1 August 2025) over three months. A median of +5.38% tells you about 637 episodes, not about one.

Data appendix

Every figure above, with the date it belongs to. Prices are closing prices; highs and lows are closing highs and lows, which run slightly narrower than the intraday 52-week range a broker screen shows. Nothing here is refreshed after publication.