What followed RSI 30 and 70 in 35 large US stocks, 1962–2026

The short answer

After RSI fell to 30 or below, 53% of 2,518 episodes beat their stock’s own drift over the next 20 sessions. After it rose to 70 or above, 45% of 4,151 did. Either way, any single episode is close to a coin toss, though the averages hold up statistically on both sides.

The bounce after an oversold reading has also mostly faded. Before 1990 it averaged between +1.47 and +2.09 points a decade; since 1990 it is +0.23 points across 1,699 episodes.

What was measured

The signal is Wilder’s RSI(14) on daily closes, read at two fixed levels: 30 or below counts as oversold, 70 or above as overbought.

An episode starts on the first day RSI crosses one of those levels. Readings within 10 sessions of each other count once, so a stock that stays oversold for weeks is not counted again and again.

Each episode is followed for 1, 5 and 20 sessions. The return over each horizon is compared with the stock’s own average return over the same horizon, worked out from the stock’s whole price history up to and including the signal day, once at least 30 such returns exist. That average is what this note calls the stock’s drift, or its usual pace. Subtracting it leaves the part of the move that the stock’s normal tendency to rise or fall does not explain. It is not a comparison with the market.

The sample is 35 large US stocks, chosen in 2026 as the large US companies people most often look up, across eight sectors, before any of these results were computed; each is followed from its first available price, 1962 to 2026: 2,518 oversold and 4,151 overbought episodes with a full 20 sessions of data after them.

Results

On both sides the tilt grows from one session to twenty, and on both it points towards a reversal.

One thing to keep in mind when reading the t-statistics: episodes bunch together. In a sell-off, many stocks turn oversold in the same week; in a rally, many turn overbought. Counting each episode as independent would overstate the evidence, so the test runs on monthly averages instead. On that basis both tilts are statistically distinguishable from zero: t +5.4 after 30 and −4.5 after 70.

Still, the averages are tiny next to the spread. The standard deviation of 20-session excess returns is 11.3 points after 30 and 9.7 after 70, while neither average reaches one point.

The oversold bounce faded after 1990

Before 1990, an oversold reading was followed by one to two points of excess over 20 sessions. Since 1990 it has been +0.23 across all 1,699 episodes.

The sample grows from 10 companies in the 1960s to 35 today, so the solid bars follow only the 10 that have been in it throughout: Boeing, Caterpillar, Chevron, Disney, Johnson & Johnson, Coca-Cola, Lockheed Martin, McDonald's, Procter & Gamble, Exxon Mobil. Their bounce fades too, so the decline is not just newer companies diluting the average. The 2020s are an unfinished decade, with 129 episodes for those 10.

What shrank was mostly the size of the bounce, not how often it came: for the same 10, 55–57% of episodes beat drift from the 1990s to the 2010s, against 53–54% before 1990.

Stocks differ widely

Individual stocks differ widely. Among the 30 with at least 30 oversold episodes, averages run from +5.10 points for Walmart to −3.85 for Netflix, and 23 of the 30 have a positive average. With that many stocks, some will look extreme by chance, so the best and worst figures probably overstate how different those companies really are.

An average can hide a skew. Tesla’s +2.39 rests on 30 episodes, only 43% of which beat drift: a few large rebounds carry it.

The overbought side spreads just as widely, from Tesla at +8.59 to Home Depot at −4.14, with 12 of 34 stocks positive.

What this does not say

This is a record of what followed in the past, not a forecast or a trading rule, and it says nothing about the next episode.

The sample is made of survivors. All 35 companies are large today, so a company that fell after an oversold reading and never recovered is not here. Leaving such companies out flatters the oversold side.

The comparison is with each stock’s own drift, not with the market: a positive number means a stock did better than its own usual pace, not that it beat an index. And no costs are deducted, whether commissions, bid–ask spreads or taxes; a tilt under one point can disappear after them.

The companies also change over time: the sample grows from 10 stocks in the 1960s to 35 in the 2020s. The section on the fading bounce controls for that.

Data and method

• Episodes and outcomes come from the site’s signal track record for RSI, rebuilt every week; this note uses its data as of 2026-10-02.
• Daily prices come from Yahoo Finance. RSI uses each stock’s trailing 250 sessions, the same value the stock pages show.
• Outcomes run close to close from the signal day; each stock’s drift uses its prices up to and including that day.
• Returns are price returns: dividends are left out of both the outcome and the drift, so the two compare like for like, except that an ex-dividend date inside a window lowers that window’s return.
• Lockheed Martin’s and Exxon Mobil’s prices run under today’s tickers through their 1995 and 1999 mergers, so their early years are Lockheed’s and Exxon’s.
• Full definitions are on the methodology page. The companion note looks at what followed a 52-week high or low.
• First edition, published 2026-10-08, with data as of 2026-10-02. The figures are fixed: a later update would appear as a new, dated edition.